Behavior-guided path planning in autonomous machine applications

Abstract

In various examples, a machine learning model (such as a deep neural network (DNN)) can be trained to use image data and / or other sensor data as input to generate two-dimensional or three-dimensional track points, vehicle positions, and / or vehicle states in world space. For example, sensor data representing the position, steering information, and / or speed of a vehicle can be collected and used to automatically generate tracks for use as ground truth data for training the DNN. Once deployed, the track points, vehicle positions, and / or vehicle states can be used by a control component (e.g., a vehicle controller) to control the vehicle through a physical environment. For example, the control component can use these outputs of the DNN to determine a vehicle-specific control profile (e.g., steering, deceleration, and / or acceleration) to control the vehicle through the physical environment.

Description

[0001] CROSS-REFERENCE TO RELATED APPLICATIONS

[0002] This application claims the benefit of U.S. Provisional Application No. 62687200, filed on June 19, 2018, which is incorporated herein by reference in its entirety.

[0003] This application is related to U.S. non-provisional application No. 16241005, filed on January 7, 2019, and the entire contents of which are incorporated herein by reference. Background Art

[0004] It is extremely difficult to design a system that can drive a vehicle autonomously without actually receiving the required level of safety supervision. Autonomous vehicles should at least be able to act as the functional equivalent of a careful driver, utilizing a perception and action system that is incredibly capable of recognizing and reacting to moving and stationary obstacles in complex environments to avoid collisions with other objects or structures in its path.

[0005] One task of an autonomous system may be to determine a recommended path for an autonomous vehicle to follow a complex environment that the vehicle may encounter. For path planning or control, some conventional systems use end-to-end trained deep neural networks (DNNs) that output object-dependent commands, such as steering angle, acceleration, deceleration, etc., that can be used by one or more components of the autonomous vehicle. For example, a DNN may use image data to compute outputs representing the steering angle, acceleration, and / or deceleration of a particular vehicle. However, because each vehicle year, brand, and / or model may have its own steering profile, acceleration profile, and / or deceleration profile, the scalability and efficacy of these end-to-end trained DNNs are significantly reduced. For example, each vehicle of a first brand and model may share the same steering angle that allows that vehicle to perform safely and operate smoothly, while each vehicle of a different brand and model may share a steering angle that is specific to that brand and model. Therefore, the steering angle output of the first brand and model may not be effectively transferred to the second brand and model (e.g., may result in unstable or unsafe vehicle control). As a result, a DNN may need to be trained individually for each brand and / or model of vehicle each year, and / or significant post-processing may be required to convert or calibrate the output from the DNN for use with different vehicle makes and models. However, by training a new DNN for each brand and / or model each year, and / or by performing extensive post-processing, significant computing resources are required in both the training of the DNNs and their deployment.

[0006] Furthermore, in some conventional systems, manual labeling or annotation may be required to effectively train a DNN. However, manual labeling or annotation may require a significant amount of time and effort - especially as the system scales. For example, when training a DNN for different environments (e.g., new terrains, new geographic locations, new scenarios, etc.) and / or with different vehicle years, makes, and / or models, new ground truth data may need to be created - requiring additional manual labels - in order to train the DNN to operate in different environments and / or different vehicles. Furthermore, because annotations or labels may be made in image space (e.g., corresponding to two-dimensional (2D) pixel coordinates), the output of the DNN may require a conversion from image space to world space (e.g., three-dimensional (3D) world coordinates) (- thereby increasing computational cost and processing time. Another consequence of manually generated labels or annotations relative to image space is that a DNN trained using these labels or annotations is less likely to generate output that represents a desired or intended path through an environment. For example, a path through image space (e.g., a path through 2D pixel locations of an image) may not correspond to a desired or intended path through 3D world space (e.g., due to difficulties in predicting curves, hills, elevation changes, and / or the like in image space), which may cause the DNN to produce noisy outputs, which may result in less stable, less smooth control of the vehicle.

[0007] summary

[0008] Embodiments of the present disclosure relate to behavior-guided path planning in autonomous machine applications. Systems and methods are disclosed that use observed behaviors (e.g., driving through a physical environment) to train one or more machine learning models to predict paths through the environment.

[0009] In contrast to conventional systems such as those described herein, the system of the present disclosure can use automatically generated ground truth data to train a machine learning model - such as a deep neural network (DNN) (e.g., a convolutional neural network CNN) - to calculate trajectory points, vehicle orientation (e.g., relative to features of the environment, such as lane markings), and / or vehicle state (e.g., relative to object manipulation, such as lane changes, turns, merges, etc.). For example, sensor data representing the vehicle's orientation, steering information, and / or speed can be collected and used to automatically generate trajectories for use as ground truth data for training the DNN. By automatically generating trajectories using sensor data as ground truth, the requirement of conventional systems for labeling and annotating new ground truth data for each new environment is eliminated, thereby increasing the scalability of the system of the present disclosure. In addition, during training, the output of the DNN can be compared to the trajectory points of the trajectory in world space, which can allow the DNN to inherently account for the conversion from 2D pixel coordinates to 2D or 3D world space coordinates without any required pre-processing or post-processing.

[0010] In deployment, the output trajectory points from the DNN can be passed to a control component (e.g., a vehicle controller) for controlling a vehicle. For example, the control component can then use the trajectory points to determine a vehicle-specific control profile (e.g., steering, braking, and acceleration) to control the vehicle through the environment based on the trajectory points. In this way, in contrast to actual control of the vehicle that can be output by the DNN of a conventional system, the trajectory points can allow the system of the present disclosure to be vehicle-independent - for example, enabling the use of vehicles of different years, makes, and / or models using the same output of the same DNN to control the vehicle according to the configuration and settings of the corresponding vehicle.

[0011] Brief Description

[0012] The proposed system and method for behavior-guided path planning in autonomous machine applications are described in detail below with reference to the accompanying drawings, wherein:

[0013] Figure 1A is a data flow diagram illustrating an example process for generating path planning information according to some embodiments of the present disclosure;

[0014] Figure 1B-1C is an illustration of one or more example machine learning models according to some embodiments of the present disclosure;

[0015] Figure 2A is an illustration of an example process for generating transformed image data according to some embodiments of the present disclosure;

[0016] Figure 2B including example visualizations of trajectories generated by one or more machine learning models after conversion to image space according to some embodiments of the present disclosure;

[0017] Figure 3 is a flow chart illustrating a method for generating trajectories from transformed image data according to some embodiments of the present disclosure;

[0018] Figure 4 is a data flow diagram illustrating an example process for training one or more machine learning models to generate path planning information according to some embodiments of the present disclosure;

[0019] Figures 5A-5C is a flow chart illustrating a method for training one or more machine learning models to generate path planning information according to some embodiments of the present disclosure;

[0020] Fig. 6A is an illustration of an exemplary autonomous vehicle according to some embodiments of the present disclosure;

[0021] Figure 6B According to some embodiments of the present disclosure Fig. 6A Examples of camera positions and fields of view for an exemplary autonomous vehicle;

[0022] Figure 6C According to some embodiments of the present disclosure Fig. 6A A block diagram of an example system architecture of an exemplary autonomous vehicle;

[0023] Fig.6D According to some embodiments of the present disclosure, one or more cloud-based servers and Fig. 6A A system diagram of exemplary autonomous vehicle communications; and

[0024] Figure 7 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure. Detailed Description

[0025] Systems and methods are disclosed that relate to behavior-guided path planning in autonomous machine applications. More specifically, the present disclosure relates to path planning for an autonomous vehicle using behavior observations represented by sensor data generated by sensors of the autonomous vehicle. Although the present disclosure may be described with respect to an example autonomous vehicle 138 (alternatively referred to herein as “vehicle 138 ” or “autonomous vehicle 138 ”), examples thereof are referred to herein as Figures 6A-6D Described herein, but not limited to. For example, the systems and methods described herein can be used by non-autonomous vehicles, semi-autonomous vehicles (e.g., in one or more advanced driver assistance systems (ADAS)), robots, warehouse vehicles, off-road vehicles, flying vessels, ships, and / or other vehicle types. In addition, although the present disclosure may be described with respect to autonomous driving, this is not limiting. For example, the systems and methods described herein can be used for robots (e.g., path planning for robots), aerial systems (e.g., path planning for drones or other aircraft), navigation systems (e.g., path planning for ships or other water containers), and / or other technical fields, such as for positioning, path planning, and / or other processes.

[0026] Trajectory Generation System

[0027] Path planning can be used to help an autonomous machine controlled by an autonomous system (e.g., an autonomous driving software stack) navigate in an environment. In an embodiment of the present disclosure, and in order to help determine the path of an autonomous vehicle, a deep neural network (DNN) (e.g., a convolutional neural network (CNN)) can be used to generate one or more trajectory points corresponding to a predicted or recommended trajectory for an autonomous vehicle. One or more output trajectory points can be generated in defined distance increments - such as 1 meter increments, 2 meter increments, etc. - and then passed to a control component (e.g., a vehicle controller) to control the autonomous vehicle. In some examples, the increment at the vehicle where the trajectory point is located can be further increased to take into account the potential reduction in accuracy of the DNN at a greater distance. The control component can then use one or more trajectory points to determine a control profile and / or a series of operations (e.g., steering, braking, and acceleration) to control the vehicle through one or more trajectory points. Generating one or more trajectory points allows the system to operate safely and autonomously without relying on the vehicle type (e.g., year, brand, and / or model), and can enable the control component (which may be different according to each vehicle year, brand, and / or model) to control the vehicle using one or more trajectory points according to the configuration and settings of the corresponding vehicle itself.

[0028] In some examples, image data used to train the DNN and image data used by the DNN during deployment may be captured by cameras with different perspectives (e.g., relative to a particular vehicle, the position and orientation of the ground plane, or other reference points). In such an example, the image data may be transformed by shifting and / or rotating the image represented by the image data to the perspective of a reference or ideal camera to generate transformed image data. In this way, the DNN can be trained on image data from the same transformed perspective. In such an example, when the DNN is deployed, another transformation may be applied to the image data from the vehicle in the deployment to transform the image data from the perspective of the vehicle's camera to the perspective of a reference or ideal camera. As a result, the DNN can be trained and deployed - using any number of different vehicles with cameras of different perspectives - using image data with the same perspective as the reference or ideal camera. This process increases the scalability of the system while removing vehicle-specific dependencies to generate machine learning models that can be deployed in any number of different vehicles.

[0029] In some embodiments, the trajectory points can be output as three-dimensional (3D) world space coordinates. By adding a third dimension (e.g., a Z, or height, dimension) to the trajectory points, the trajectory generated by the machine learning model can be more accurate and, therefore, more effective in path planning for the vehicle. For example, using only two-dimensional (2D) points, a path around a lateral curve or turn that includes a vertical curve (e.g., a hill) may be different than a flat surface with the same lateral curve or turn on it. This may result in less accurate trajectory points from the machine learning model (e.g., a path through a curve may be offset to the left or right, driving closer to a lane marking or boundary than the center of the lane). However, by training the machine learning model using three dimensions, the machine learning model can learn to account for these vertical curves and can therefore produce consistent trajectory point outputs for vertically curved and flat surfaces (e.g., because the turns may be essentially the same and only the high ground may change).

[0030] Now refer to Figure 1, Figure 1A 1 is a data flow diagram illustrating an example process 100 for generating path planning information according to some embodiments of the present disclosure. At a high level, the process 100 may include one or more machine learning models 108 that receive one or more inputs 102 and generate one or more outputs 110. The one or more inputs 102 may include image data, such as image data generated by a camera of one or more autonomous vehicles (e.g., vehicle 138 as described herein), in addition to sensor data 106 in some embodiments (e.g., LIDAR data from one or more LIDAR sensors, RADAR data from one or more RADAR sensors, audio data from one or more microphones, SONAR data from one or more SONAR sensors, etc.). The one or more machine learning models 108 may be trained to generate trajectory data 112 (e.g., one or more trajectory points in a 2D or 3D world space), vehicle orientation 114 (e.g., relative to one or more features of the environment, such as lane markings, lanes, road boundaries, etc.) and / or vehicle state 116 (e.g., regarding vehicle maneuvers, such as lane changes, turns, lane splits, etc.). These outputs 110 may be used by one or more control components 118 of the autonomous vehicle (e.g., one or more controllers 636, ADAS systems 638, one or more SOCs 604, and / or other components of the autonomous vehicle 138) to help the vehicle navigate (e.g., path planning) an environment.

[0031] The one or more inputs 102 may include image data 104, sensor data 106, and / or other types of data (e.g., map data). The image data 104 may include data representing an image of a field of view of one or more cameras of the vehicle (such as one or more stereo cameras 868, one or more wide view cameras 870 (e.g., fisheye cameras), one or more infrared cameras 872, one or more surround cameras 874 (e.g., 360 degree cameras), one or more long-range and / or mid-range cameras 898, and / or other camera types of the vehicle 138). In some examples, the image data 104 may be captured by a single camera having a forward, substantially centered field of view relative to the horizontal axis (e.g., from left to right) of the vehicle 138. In a non-limiting embodiment, one or more forward-facing cameras (e.g., center- or near-center-mounted cameras) may be used, such as the wide view camera 670, the surround camera 674, the stereo camera 668, and / or the long-range or mid-range camera 698. Image data 104 captured from this perspective may be useful for computer vision and / or perception when navigating (e.g., within a lane, through a lane change, through a turn, through an intersection, etc.) because the forward-facing camera may include a field of view (e.g., Figure 6B The field of view of the forward-facing stereo camera 668 and / or the wide-view camera 670 may include the current lane of travel of the vehicle 138, one or more adjacent lanes of travel of the vehicle 138, and / or the boundaries of the driving surface. In some examples, more than one camera or other sensor (e.g., a LIDAR sensor, a RADAR sensor, etc.) may be used to combine multiple fields of view (e.g., Figure 6B of the remote camera 698, the forward stereo camera 668, and / or the forward wide field of view camera 670).

[0032] In some examples, the image data 104 may be captured in one format (e.g., RCCB, RCCC, RBGC, etc.) and then converted (e.g., during preprocessing of the image data 104) to another format. In some other examples, the image data 104 may be provided as input to a sensor data preprocessor (not shown) to generate preprocessed image data (also referred to herein as image data 104). Many types of images or formats may be used as input, for example, compressed images such as Joint Photographic Experts Group (JPEG), Red Green Blue (RGB), or Luminance / Chrominance (YUV) formats, compressed images as frames derived from compressed video formats (e.g., H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC)), such as raw images derived from Red Clear Blue (RCCB), Red Clear (RCCC), or other types of imaging sensors. Note that different formats and / or resolutions may be used to train one or more machine learning models 108 rather than for inference (e.g., during deployment of one or more machine learning models 108 in an autonomous vehicle 138).

[0033] The sensor data preprocessor can use image data representing one or more images (or other data representations) and load the sensor data into memory in the form of a multidimensional array / matrix (or referred to as a tensor, or more specifically referred to as an input tensor in some examples). The array size can be calculated and / or represented as W×H×C, where W represents the image width in pixels, H represents the height in pixels, and C represents the number of color channels. Without loss of generality, other types and orders of input image components are also possible. In addition, the batch size B can be used as a dimension (e.g., an additional fourth dimension) when batching is used. Batching can be used for training and / or reasoning. Therefore, the input tensor can represent an array of dimensions W×H×C×B. Any ordering of dimensions is possible, which may depend on the specific hardware and software used to implement the sensor data preprocessor. The ordering can be selected to maximize the training and / or reasoning performance of one or more machine learning models 108.

[0034] In some embodiments, the sensor data preprocessor may employ a preprocessed image transmission path to process one or more raw images acquired by one or more sensors (e.g., one or more cameras) and included in the image data 104 to generate preprocessed image data, which may represent one or more input layers of the one or more machine learning models 108 (e.g., Figure 1BThe convolution stream 120 of FIG. 120 may be used to process one or more input images of the convolution stream 120 of FIG. 120 . An example of a suitable pre-processing image transfer path may use a raw RCCB Bayer (e.g., 1 channel) type image from a sensor and convert the image to an RCB (e.g., 3 channel) planar image format stored with a fixed precision (e.g., 16 bits per channel). The pre-processing image transfer path may include decompression, noise reduction, demosaicing, white balancing, histogram calculation, and / or adaptive global tone mapping (e.g., in this order or in an alternate order).

[0035] In the case where the sensor data preprocessor employs noise reduction, it may include bilateral denoising in the Bayer domain. In the case where the sensor data preprocessor employs demosaicing, it may include bilinear interpolation. In the case where the sensor data preprocessor employs histogram calculation, it may involve calculating a histogram of the C channel and may be combined with decompression or noise reduction in some examples. In the case where the sensor data preprocessor employs adaptive global tone mapping, it may include performing an adaptive gamma log profile transform. This may include calculating a histogram, obtaining mid-tone levels and / or estimating the maximum brightness using mid-tone levels.

[0036] In some examples, the sensor data 106 may include sensor data generated by any number of sensors, such as one or more LIDAR sensors 664, one or more RADAR sensors 660, one or more ultrasonic sensors 662, one or more microphones 696, and / or other sensor types. The sensor data 106 may represent the field of view and / or sensing field of a sensor (e.g., one or more LIDAR sensors 664, one or more RADAR sensors 660, etc.), and / or may represent the environmental perception of one or more sensors (e.g., one or more microphones 696). Sensors such as image sensors (e.g., cameras), LIDAR sensors, RADAR sensors, SONAR sensors, ultrasonic sensors, etc. may be referred to herein as perception sensors or perception sensor devices, and sensor data generated by perception (herein may be referred to as perception sensor data). In some examples, an instance or representation of the sensor data 106 may be represented by an image captured by an image sensor, a depth map generated by a LIDAR sensor, etc. LIDAR data, SONAR data, RADAR data, and / or other sensor data types may be related or associated with image data generated by one or more image sensors. For example, image data representing one or more images may be updated to include data related to a LIDAR sensor, a SONAR sensor, a RADAR sensor, etc., so that the sensor data used for training and / or input to the DNN may be more informative or detailed than image data alone. In this way, the DNN may learn to generate trajectories in a 3D or 2D world space using this additional information from any number of perception sensors.

[0037] In embodiments using sensor data 106, the sensor may be calibrated so that the sensor data 106 is associated with pixel coordinates in the image data 104. The sensor data preprocessor may also perform preprocessing on the sensor data 106, the sensor data 106 being similar to the preprocessed data described herein with respect to the image data 104. In some embodiments, such as where the sensor data 106 indicates depth (e.g., RADAR data, LIDAR data, etc.), depth values ​​may be associated with pixel coordinates in the image data 104 and then used as additional (or alternative, in some examples) inputs 102 to the machine learning model 108. For example, as determined from the sensor data 106, one or more pixels may have additional values ​​associated with their depth representation.

[0038] In some examples, the image data 104 and / or sensor data 106 may be transformed, calibrated, and / or augmented to a reference or ideal camera and / or sensor, respectively. For example, to increase the scalability of the system and remove vehicle type dependencies, one or more machine learning models 108 may be trained and deployed to use image data 104 and / or sensor data 106 from any of the same reference or ideal camera and / or sensor perspectives (e.g., position—left to right, and top to bottom—and / or rotation). Thus, camera parameters (e.g., intrinsic and / or extrinsic) may be used to determine the perspective of one or more cameras that captured the training image data, and the training image data may be transformed or augmented to the perspective of a reference or ideal camera (e.g., a camera that is centered on the vehicle and mounted at a location such as a bumper or the front of the vehicle, from left to right). In deployment, camera parameters may be used to determine the perspective of one or more cameras that captured the image data 104, and the image data 104 may be transformed or augmented to the perspective of a reference or ideal camera. As a result, one or more machine learning models 108 are trained and deployed using image data from essentially the same perspective. Similarly, for sensors, sensor parameters can be used to transform or augment the training sensor data and sensor data 106 to the perspective of a reference or ideal sensor. Figure 2A and Figure 3 The transformation and / or augmentation of image data 104 and / or sensor data 106 is described in greater detail.

[0039] In some embodiments, the input 102 may include other data types, such as map data. The map data may be used by one or more machine learning models 108 to generate outputs 110, such as vehicle states 116 and / or trajectories 112. For example, the map data may include low-resolution map data (e.g., screenshots of a 2D map application with or without guidance). The low-resolution map data may include basic geometry of roads and / or intersections, for example, without additional information such as lane markings, number of lanes, location of sidewalks, street lights, stop signs, etc. In other words, the map data may be less data-intensive than map data representing an HD map (e.g., the HD map 622 and / or HD maps described herein and relied upon by conventional systems) and used by one or more machine learning models 108 as only additional data points when calculating the output 110.

[0040] In some examples, the map data may include screenshots or images (or data representative thereof) depicting a current lane of the vehicle, a destination lane of the vehicle, a representation of the vehicle itself, and / or a path of the vehicle through lane changes. In some examples, the path of the vehicle used for training map data may be automatically generated during a human-driven portion of vehicle operation (e.g., the path is populated on a map while the vehicle is controlled through the environment).

[0041] In an example, the map data may include commands such as "At the next intersection, turn right," and the one or more machine learning models 108 may use this information to generate trajectory data 112 and / or vehicle states 116 corresponding to staying in the current lane, or switching lanes to reach the rightmost lane, and then turning right at the next intersection. Similarly, for a lane split, the command may represent staying on the left side during a lane split, and as the vehicle is maneuvered to the left during the lane split, and the one or more machine learning models may use this information to generate trajectory data 112 and / or vehicle states 116. In any example, the map data may be automatically generated (e.g., during a person driving a car) and / or may be generated by manual labeling.

[0042] In some examples, the map data may be similar to and / or may be used (e.g., for training and / or inference) in a manner similar to the map information described in U.S. Non-Provisional Application No. 16,241,005, filed on January 7, 2019. The entire contents of which are incorporated herein by reference.

[0043] One or more machine learning models 108 may use one or more images or other data representations or instances (e.g., LIDAR data, RADAR data, SONAR data, ultrasound data, etc.) represented by sensor data 106 as input to generate one or more outputs (110). In one non-limiting example, one or more machine learning models 108 may use one or more images represented by sensor data 106 as input (e.g., after pre-processing to generate trajectory data 112, vehicle position 114, and / or vehicle state 116). Although examples using neural networks, and in particular convolutional neural networks, are described herein, one or more machine learning models 108 (e.g., regarding Figure 1B and Figure 1C ), which is not limiting. For example, but not limited to, the one or more machine learning models 108 described herein may include any type of machine learning model, such as using linear regression, logistic regression, decision trees, support vector machines (SVM), k-nearest neighbor (Knn), K means clustering, random forest, dimensionality reduction algorithm, gradient boosting algorithm, neural network (e.g., autoencoder, convolutional, recurrent, perceptron, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid state machine, etc.) machine learning model, and / or other types of machine learning models.

[0044] One or more outputs 110 may include trajectory data 112, vehicle position 114, and / or vehicle state 116. The trajectory data 112 may include one or more trajectory points representing 2D or 3D world space coordinates (e.g., coordinates in a physical environment, rather than pixel coordinates of conventional methods). For example, each row of trajectory data 112 may be output as a two-cell array for each trajectory point in an embodiment generating 2D coordinates (e.g., a first cell for x position and a second cell for y position), and may be output as a three-cell array for each trajectory point in an embodiment generating 3D coordinates (e.g., a first cell for x position, a second cell for y position, and a third cell for z position). Depending on the embodiment, any number of trajectory points may be generated for a single trajectory. For example, in some examples, a single trajectory point may be generated, while in other examples, forty to sixty trajectory points may be generated. This is not limiting, and one or more machine learning models 108 may be trained to generate any number of trajectory points without departing from the scope of the present disclosure. The trajectory data 112 output by the one or more machine learning models 108 may include a number of rows corresponding to the number of trajectory points calculated by training the one or more machine learning models. As a non-limiting example, where the one or more machine learning models 108 are trained to generate forty trajectory points in 3D world coordinates, the trajectory data 112 (for each iteration of the one or more inputs 102) may include a 3×40 output array.

[0045] The location represented by the 2D or 3D coordinates may represent the location where the origin (e.g., 2D or 3D location on the vehicle) of the vehicle (e.g., vehicle 138) should navigate (e.g., the origin of the vehicle is substantially aligned with or overlaps the 2D or 3D coordinates). According to a specific embodiment, the origin of the vehicle (e.g., [0,0] for 2D and [0,0,0] for 3D) may be an arbitrary point that depends on the specific embodiment (e.g., based on the origin used to generate training trajectory data for one or more machine learning models). For example, the origin may be located at the lateral center of an axle of the vehicle (e.g., the rear axle, the front axle, the center axle in a vehicle with more than two axles, etc.), the longitudinal center of the vehicle, the vertical center of the vehicle, the front of the vehicle, the rear of the vehicle, a side of the vehicle, another location of the vehicle, and / or a combination thereof.

[0046] As a non-limiting example, assuming a first track point (e.g., [x, y, z], expressed in feet, with positive values ​​indicating forward, right, or up, and negative values ​​indicating backward, left, or down) is at [1.4, -0.7, 0], the origin of the vehicle can be controlled to be 1.4 feet forward, 0.7 feet left, and 0 feet in the vertical or z direction (e.g., flat ground). Each successive track point can include a value corresponding to the origin of a particular track (e.g., if the second track point is at [3.1, -0.9, 0], this would be relative to the origin, or at the beginning of the vehicle's path), or each successive track point can correspond to a difference from a previous (e.g., immediately preceding) track point (e.g., such that the previous track point is zeroed and the next track point is relative to the previous zeroed track point). Track points can be expressed in polar coordinates, Cartesian coordinates, and / or other forms. Although discussed herein as being expressed in feet or meters, this is not limiting. For example, the measurements associated with the track points can be in any units, such as SI, metric, or US, to name a few examples. For example, depending on the embodiment, millimeters, centimeters, inches, feet, meters, and / or any other units of measurement may be used without departing from the scope of the present disclosure.

[0047] Track points may be calculated in increments of world space (e.g., in the x-direction or forward direction), such as one meter, two meters, three meters, ten meters, etc. For example, the increment size may be determined or optimized during training to generate track points that are effective for the vehicle (e.g., resulting in safe, smooth control of the vehicle) while not requiring the calculation of so many track points that the processing speed is reduced to less than real-time operation or uses unnecessary computing resources. Thus, without limitation, an increment size of one to two meters provides accurate and usable results during testing that allows process 100 to execute in real-time without hindering the real-time operation of any of the many other processes being executed simultaneously by the vehicle (e.g., processes executed by any number of layers of the autonomous driving software stack, such as a planning layer, a control layer, an actuation layer, an obstacle avoidance layer, etc.).

[0048] The increment size, combined with the number of trajectory points to be calculated, can determine how far in world space the trajectory extends from the vehicle (or from the beginning of the vehicle's path). For example, in a non-limiting example where the increment size is one meter and fifty trajectory points are calculated, the predicted or recommended trajectory calculated by one or more machine learning models 108 may extend fifty meters. The length of the desired trajectory can be a factor used to determine the increment size and / or number of trajectory points calculated by one or more machine learning models 108. Thus, the length of the trajectory can be, but is not limited to, one meter, two meters, twenty meters, forty meters, sixty meters, one hundred and eighty meters, etc. without departing from the scope of the present invention.

[0049] In some examples, because the accuracy of the predicted trajectory points decreases the farther they are from the vehicle (or from the origin of the vehicle), the incremental size can increase at greater distances. For non-limiting examples, the incremental size can be a first size for the first trajectory point, a second size for the subsequent trajectory points, a third size for the subsequent trajectory points, a fourth size for the subsequent trajectory points, and so on, where the size increases from the first to the fourth. The change in size can be gradual (e.g., it can increase steadily, such as by an x ​​amount per trajectory point, it can increase at intervals, such as by every x trajectory points, etc.), or it can be more clearly depicted (e.g., the trajectory points in the first half have a first incremental size, the trajectory points in the second half have a second incremental size that is larger than the first incremental size, etc.)

[0050] In some examples, trajectory data 112 may include a parameterized representation of a path, such as parameters of a polygonal approximation of the path, without departing from the scope of this disclosure.

[0051] In some exemplary embodiments, a vehicle (e.g., vehicle 138) may include a positioning system (e.g., using HD map 622, one or more GNSS sensors 658, one or more components of a perception layer of an autonomous driving software stack, etc.) that is capable of calculating geographic location with high accuracy (e.g., GNSS alone may only be accurate to within 30 meters, a positioning system may be accurate to within a centimeter or less). In such an embodiment, the positioning information may be tied to the origin of the vehicle, and the track points may be converted to GNSS coordinates before or after use by control component 118. For example, one or more control components 118 may use the GNSS coordinates to navigate the vehicle to the GNSS coordinates and / or the GNSS coordinates may be used by one or more other components of the vehicle and / or the autonomous driving software stack (e.g., a planning layer, a control layer, one or more control components of the perception layer, and / or a component of another layer of the stack).

[0052] According to an embodiment, another output 110 (vehicle orientation 114) may represent the orientation of the vehicle relative to a lane, a road boundary, another vehicle, etc. For example, one or more machine learning models 108 may be trained to output an orientation relative to a lane (e.g., inside a lane, outside a lane, between lanes), at an angle (e.g., a specific angle (such as degrees) or a value representing degrees) relative to a lane marking, etc. The vehicle orientation 114 may help inform one or more control components 118 about the current direction of the vehicle to help determine how to control the vehicle to traverse a trajectory point.

[0053] Another output 110 may be a vehicle state 116. As described herein, the vehicle state 116 may be a state regarding a vehicle maneuver, such as a turn, a lane change, a lane split, a lane merge, and / or other vehicle maneuvers. In such an example, the input 102, output 110, and / or training process of one or more machine learning models 108 may be at least partially similar to the disclosure of U.S. Non-Provisional Application No. 16,241,005, filed on January 7, 2019, the entire contents of which are incorporated herein by reference. The vehicle state 116 may inform one or more control components 118 how to control the vehicle, such as preparing to re-enter a lane keeping mode after a lane change mode.

[0054] In some examples, one or more machine learning models 108 may be trained to output validity scores for some or all track points. In such examples, one or more machine learning models 108 may generate a confidence score that represents the likelihood that one or more track points are accurate. For example, in the case where a track point is outside the field of view of the camera, the confidence score may be low and the track point may be marked as invalid. A threshold confidence score may be used to determine whether a track point is valid or invalid. When invalid, the track point may be discarded or ignored, and a track point from the next or previous iteration of one or more machine learning models 108 may be used instead, or the track point may be skipped for the next track point of the same track starting from the current iteration.

[0055] The one or more control components 118 can use the output of the one or more machine learning models 108 to perform one or more operations. The one or more operations can include controlling the vehicle toward or toward a trajectory point represented by the trajectory data 112 (and in some embodiments can use the vehicle position 114 and / or vehicle state 116 for context). Because the one or more machine learning models 108 can be an end-to-end system - meaning that the input data is received as input and the output can be used directly by the vehicle - the one or more control components 118 can receive and / or use the one or more outputs 110 without any post-processing. By reducing the amount of post-processing to zero, the process 100 can be performed efficiently, accurately, and in real time without unnecessarily consuming computing resources (as opposed to conventional systems, as described herein).

[0056] However, in some examples, post-processing such as temporal smoothing, spatial smoothing, GNNS conversion of the trajectory data 112 (e.g., as described herein), and / or other post-processing may be performed. For example, trajectory data 112 from a current iteration of one or more machine learning models 108 may use temporal smoothing to weight the trajectory data 112 against trajectory data 112 from any or all previous iterations. As another example, some trajectory data 112 from the current iteration may be weighed against other trajectory data 112 from the current iteration to perform spatial smoothing. In such an example, where a threshold difference in any direction (e.g., x, y, or z) is determined between two or more points (e.g., consecutive points), a calculation may be performed to adjust the value (e.g., by averaging, using a standard deviation, etc.). The result of spatial smoothing and / or temporal smoothing may be a less noisy, more fluid or smooth output 110.

[0057] Reference now Figure 1B , Figure 1B is an illustration of one or more example machine learning models 108A according to some embodiments of the present disclosure. Figure 1B The one or more machine learning models 108A may be an example of one or more machine learning models 108 that may be used in process 100. However, Figure 1B The one or more machine learning models 108A are not limiting, and the one or more machine learning models 108 may include more Figure 1B One or more machine learning models 108A may include or be referred to as a convolutional neural network (CNN), and may therefore be referred to herein alternatively as a convolutional neural network 108A, a convolutional network 108A, or a CNN 108A.

[0058] CNN 108A may use image data 104 and / or sensor data 106 (with or without any pre-processing) as input. For example, CNN 108A may use image data 104 (as shown in images 104A, 104B, and 104C) as input. Image data 104 may represent images acquired by one or more cameras (e.g., Figures 6A-6C108A). For example, image data 104 may represent the field of view of one or more cameras. More specifically, image data 104 may represent individual images generated by one or more cameras, and image data 104 representing one or more individual images may be input to CNN 108A at each iteration of CNN 108A. In some embodiments, in addition to or as an alternative to image data 104, sensor data 106 may be input to CNN 108A. Sensor data 106 may represent a perspective view of a physical environment observed by one or more sensors, such as a LIDAR sensor, one or more RADAR sensors, one or more microphones, one or more SONAR sensors, etc.

[0059] The image data 104 may be input to one or more convolutional layers 122 (e.g., convolutional layer 122A) of the CNN 108A. The convolutional stream 120 may include any number of layers 122, such as layers 122A-122C. The sensor data 106 may be input to one or more convolutional layers 126 (e.g., convolutional layer 126A) of the CNN 108A. The convolutional stream 124 may include any number of layers 126, such as layers 126A-126C. In some examples, the image data 104 and / or sensor data from each different camera and / or each different sensor, respectively, may be associated with its own convolutional stream (e.g., image data 104 from a first camera has a first convolutional stream, image data 104 from a second camera has a second convolutional stream, sensor data 106 from a first sensor has a third convolutional stream, and so on). However, in some examples, some sensor data 106 and / or image data 104 may not be associated with any convolutional layer. In such examples, one or more other layer types may be used for at least some of sensor data 106 and / or image data 104 without departing from the scope of this disclosure.

[0060] One or more of the layers 122, 126 may include an input layer. The one or more input layers may hold values ​​associated with the image data 104 and / or the sensor data 106. For example, with respect to the image data 104, the one or more input layers may hold values ​​representing raw pixel values ​​of one or more images as a volume (e.g., width W, height H, and color channels C (e.g., RGB), such as 32×32×3), and / or a batch size B.

[0061] One or more layers 122, 126 may include a convolutional layer. A convolutional layer may compute the output of neurons connected to local regions in an input layer (e.g., an input layer), with each neuron computing a dot product between their weights and the small regions they are connected to in the input volume. The result of a convolutional layer may be another volume with one dimension based on the number of filters applied (e.g., width, height, and number of filters, such as 32×32×12 if 12 is the number of filters).

[0062] One or more of the layers 122, 126 may include a rectified linear unit (ReLU) layer. For example, one or more ReLU layers may apply an element-wise activation function (e.g., max(0, x)) thresholded to zero. The final volume of the ReLU layer may be the same as the input volume of the ReLU layer.

[0063] One or more of the layers 122, 126 may include a pooling layer. The pooling layer may perform a downsampling operation along the spatial dimensions (e.g., height and width), which may produce a smaller volume than the input of the pooling layer (e.g., 16×16×12 from a 32×32×12 input volume). In some examples, CNN 108A may not include any pooling layers. In such examples, a strided convolution layer may be used instead of a pooling layer.

[0064] One or more of the layers 122, 126 may include fully connected layers. Each neuron in the one or more fully connected layers may be connected to each neuron in the previous volume. The fully connected layers may calculate class scores, and the resulting volume may be 1×1×the number of classes. In some examples, one or more convolutional streams 120, 124 may include fully connected layers, while in other examples, the fully connected layers of CNN 108A may be separate fully connected layers from one or more convolutional streams 120, 124 (e.g., as described in

[0047] ). Figure 1C As shown, one or more fully connected layers may be part of one or more output layers 128).

[0065] Although input layers, convolutional layers, pooling layers, ReLU layers, and fully connected layers are discussed herein with respect to one or more convolutional layers 122, 126, this is not limiting. For example, additional or alternative layers 122, 126, such as normalization layers, SoftMax layers, and / or other types of layers, may be used in one or more convolutional flows 120, 124.

[0066] Depending on the embodiment, different orders and numbers of layers 122, 126 of the CNN 108A may be used. For example, for a first vehicle, there may be a first order and number of layers 122, 126, while for a second vehicle there may be a different order and number of layers 122, 126; for a first camera, there may be an order and number of layers 122 and 126 that is different from the order and number of layers for a second camera. In other words, the order and number of layers 122, 126 of the convolutional streams 120, 124 and / or one or more output layers 128 of the CNN 108A are not limited to any one architecture.

[0067] In addition, for example, some of the layers 122, 126 may include parameters (e.g., weights and / or biases), such as layers of the convolutional streams 120, 124 and / or one or more output layers 128, while other layers may not (e.g., ReLU layers and pooling layers). In some examples, the parameters may be learned by the convolutional streams 120, 124 and / or one or more machine learning models 108A during training. In addition, some of the layers 122, 126 (e.g., convolutional layers 122, 126, one or more output layers 128, and pooling layers (as part of the convolutional streams 120, 124)) may include additional hyperparameters (e.g., learning rate, stride, epochs, kernel size, number of filters, type of pooling used for merging layers, etc.), while other layers (e.g., ReLU layers) may not include these. Various activation functions may be used, including but not limited to ReLU, leaky ReLU, sigmoid, hyperbolic tangent (tanh), exponential linear unit (ELU), etc. The parameters, hyperparameters and / or activation functions are not limited and may vary depending on the embodiment.

[0068] The output of one or more convolution streams 120, 124 may be input to one or more output layers 128 (which may be one or more fully connected layers in some examples) of the CNN 108A. In addition to the output of one or more convolution streams 120, 124, in examples using vehicle orientation 114 and / or vehicle state 116, the output of one or more layers may be trained to compute vehicle orientation 114 and / or vehicle state 116 may be provided as input to one or more output layers 128 (e.g., one or more fully connected layers as described herein).

[0069] The output 110 of the CNN 108A may include trajectory data 112, vehicle position 114, vehicle state 116, and / or other output types. The trajectory data 112 may represent a recommended trajectory 142. However, in some examples, the output 110 may include control data for following the recommended trajectory 142 (e.g., for controlling the vehicle 138 according to the recommended trajectory 142, such as steering angle, acceleration, deceleration, etc.). In some examples, the trajectory data 112 may include one or more trajectory points 140 along the recommended vehicle trajectory 142 (e.g., as 2D or 3D coordinates in world space, and / or 2D coordinates in image space). For example, one or more machine learning models 108A may output only a single trajectory point 140 (e.g., the next trajectory point of the vehicle 138 in a sequence of discrete trajectory steps). In other examples, more than one trajectory point 140 may be output. As another example, the entire trajectory may be output, which may be extrapolated from two or more trajectory points 140. In some examples, the recommended trajectory 142 may be output as a radius of the recommended vehicle trajectory 142, or may be output as a reverse radius of the recommended trajectory 142 (e.g., including turns, lane changes, lane splits, lane adjustments, and / or curves in the recommended trajectory 142). The recommended trajectory 142 and / or trajectory points 140 thereon may be used by one or more control components 118 to control the vehicle 138 through the physical environment.

[0070] Reference now Figure 1C , Figure 1C is an illustration of another exemplary machine learning model 108B according to some embodiments of the present disclosure. Figure 1C One or more machine learning models 108B may be an example of a machine learning model 108 that may be used in process 100. However, Figure 1C The one or more machine learning models 108B are not limiting, and the one or more machine learning models 108 may include Figure 1C One or more machine learning models 108B may include or be referred to as a convolutional neural network (CNN), and may therefore be referred to herein alternatively as a convolutional neural network 108B, a convolutional network 108B, or a CNN 108B.

[0071] The CNN 108B may receive one or more inputs 102, which may include image data 104 and / or sensor data 106 from one or more sensor types. The one or more inputs 102 may be applied to one or more input layers 144. In a non-limiting example, there may be three input layers 144, and one or more layers may receive image data 104 in YUV, RGB, JPEG, and / or other formats at the one or more input layers 144. Depending on the embodiment, without limitation, the one or more input layers 144 may have a spatial resolution of 66×200, or may have a smaller or larger spatial resolution.

[0072] CNN 108B may include one or more normalization layers 146. The normalization layer may represent one or more normalization functions, such as batch normalization, local contrast normalization, a simplified whitening layer, local response normalization, and / or another normalization function.

[0073] CNN 108B may include one or more convolutional streams 148, which include a plurality of different layers, such as those described herein with respect to Figure 1B In a non-limiting example, one or more convolutional streams 148 may include multiple convolutional layers that reduce the spatial resolution throughout the convolutional stream 148. For example, but not limited to, Figure 1C As shown, the convolutional layer can reduce the spatial resolution from 66×200 to 31×98, 14×47, 5×22, 3×20, and 1×18. To reduce the spatial resolution, filters or kernels of any size in a variety of sizes can be used, such as a 5×5 kernel, a 3×3 kernel, etc. To reduce the spatial resolution to the final spatial resolution of the convolution stream 148 (e.g., 1×18, as shown in FIG. 1 ), the convolutional layer can reduce the spatial resolution from 66×200 to 31×98, 14×47, 5×22, 3×20, and 1×18. To reduce the spatial resolution, filters or kernels of any size in a variety of sizes can be used, such as a 5×5 kernel, a 3×3 kernel, etc. Figure 1C As shown in FIG. 1 , any number of convolutional layers may be present. Additionally, to reduce the spatial resolution from a first spatial resolution (e.g., 66×200) to a second spatial resolution (e.g., 31×98), one, three, twenty-four, thirty-six, forty-eight, sixty-four, and / or a different number of layers may be used in the convolutional stream 148.

[0074] CNN 108B may include one or more fully connected layers 150. One or more fully connected layers 150 may receive the output of one or more convolution streams 148, or, depending on the embodiment, one or more layers may exist between the output of one or more convolution streams 148 and one or more fully connected layers 150. One or more fully connected layers 150 may also be referred to here as one or more output layers of CNN 108B, and generate output 110. The final fully connected layer may include a plurality of neurons corresponding to the output array, as discussed herein. For example, in the case where CNN 108B calculates forty trajectory points in 3D world space, the final fully connected layer 150 may include 120 neurons (e.g., 40 trajectory points each with x, y, and z values, so 40×3). In other examples, the number of output neurons may be different depending on the number of trajectory points to be calculated by CNN 108B and whether the trajectory points are calculated in 2D or 3D space.

[0075] Although CNN 108B is described with respect to one or more normalization layers 146, one or more convolutional layers, one or more input layers 144, and one or more fully connected layers 150, this is not limiting. According to embodiments, CNN 108B may include additional or alternative layer types, such as but not limited to those described herein (e.g., with respect to Figure 1B ). In addition, the filter size, number of filters, number of layers, hyperparameters, functionality and / or other features and functions of CNN108B may vary depending on the embodiment without departing from the scope of the present disclosure.

[0076] Reference now Figure 2A , Figure 2Ais an illustration of an example process 200 for generating transformed image data according to some embodiments of the present disclosure. The process 200 may be used to train one or more machine learning models 108 to learn accurate trajectories even when the vehicle 138 is off-center and / or off-orientation. For example, because a driver may not want to be off-center or off-orientation during data collection, a viewpoint transformation may be used to simulate that effect so that one or more machine learning models 108 learn to generate accurate trajectories that can be used even when the vehicle 138 is off-center and / or off-orientation. Additionally, the process 200 may be used to account for the position, orientation, lens distortion, and / or other variables of different cameras or sensors and their associated vehicles that may capture image data 104 and / or sensor data 106 during training. For example, cameras and / or sensors may be positioned differently on different vehicle types (e.g., left or right relative to the center of the ground plane, up or down relative to the ground plane, etc.), may be at different rotations (e.g., tilted left, tilted right, tilted down, tilted up, etc.), and / or may have different lens distortions (or other camera parameters), each of which may affect the perspective of the camera and / or sensor.

[0077] Process 200 may include using an original image 202 (e.g., an image captured by a camera) and shifting (e.g., as shown in shifted image 204A) and / or rotating (e.g., as shown in rotated image 204B) the original image 202 to generate a transformed (or augmented) image 206. For example, lane markings 208 and trees 210 from original image 202 may be shifted (e.g., to the left) and / or rotated (e.g., in a clockwise direction) in shifted image 204A to produce transformed image 206, which is displayed in FIG. Figure 2A 204A and 204B are included in the illustration of lane markings 208 and trees 210 that are shifted and rotated relative to their positions in the original image 202. Although shown as two separate processes and including two separate images (e.g., 204A and 204B), the transformations may be applied as a single process (e.g., using a single transformation) that moves and / or rotates the original image 202 to generate the transformed image 206. As such, the shifted image 204A and the rotated image 204B may be used for visualization and example purposes only, and may not actually be generated during execution of the process 200.

[0078] In some examples, the determination of the transformation to be applied can be based on an ideal or reference camera or sensor. For example, during data capture (e.g., training image data 404 and / or Figure 4Training sensor data 406) may be generated for one or more machine learning models 108 using any number of different cameras and / or sensors from any number of different vehicles. However, the image data and / or sensor data may not be from the same perspective for each camera and / or each sensor. For example, as described herein, the position, orientation, lens distortion, and / or other camera or sensor parameters (intrinsic and / or extrinsic) may be different for each different camera, sensor, and / or vehicle. Thus, in order to train one or more machine learning models 108 using data from a consistent perspective, the image data and / or sensor data may be transformed (e.g., shifted, rotated, etc.) to the perspective of an ideal or reference camera (e.g., a virtual or imaginary camera having a set of parameters). The transformed image data and / or transformed sensor data may be used as training image data 404 and / or training sensor data 406 to train one or more machine learning models 108 ( Figure 4 ).

[0079] Furthermore, during deployment of one or more machine learning models 108 on any number of different vehicles, which may use any number of different cameras and / or sensors, the captured image data 104 and / or sensor data 106 may be similarly transformed to the perspective of an ideal or reference camera or sensor. For example, the image data 104 and / or sensor data 106 may be shifted, rotated, and / or otherwise transformed (or augmented) to generate transformed image data and / or transformed sensor data from the perspective of an ideal or reference camera. In this way, one or more machine learning models 108 may be trained and deployed using image data and / or sensor data from the same one or more perspectives. This may result in more accurate results for one or more machine learning models 108, particularly in embodiments where one or more machine learning models 108 are end-to-end systems whose outputs 110 may be used directly by a vehicle 138 (e.g., without any post-processing by one or more control components 118).

[0080] Figure 2B Included are example visualizations of trajectories generated by one or more machine learning models after conversion to image space according to some embodiments of the present disclosure. Figure 2BA first visualization 212A, a second visualization 212B, and a third visualization 212C (collectively referred to herein as “visualizations 212”) are included. Visualizations 212 may include representations of trajectory data 112 generated by one or more machine learning models 108 in 2D image space and associated with an image represented by image data 104. For example, because one or more machine learning models 108 may generate trajectory points in 2D or 3D world space, in order to visualize these trajectories, it may be necessary to convert trajectory data 112 to 2D image space coordinates. In this way, Figure 2B Three example visualizations are included of trajectories 214A, 214B, and 214C (converted from trajectory data 112 output by one or more machine learning models 108) extending from vehicle 138 to a physical environment (e.g., along a driving surface). Visualizations 212 may be used to determine key performance indicators (KPIs), for display during testing, for display during deployment (e.g., on an HMI display 634 of vehicle 138), and / or for other purposes.

[0081] Reference now Figure 3 , each block of the method 300 described herein comprises a computing process that can be performed using any combination of hardware, firmware, and / or software. For example, the various functions can be performed by a processor executing instructions stored in a memory. The method 300 can also be embodied as computer usable instructions stored on a computer storage medium. The method 300 can be provided by a stand-alone application, a service or a hosted service (standalone or in combination with another hosted service), or a plug-in to another product, to name a few examples. In addition, as an example, the method 300 described herein is described with respect to Figure 1A The process of 100 and Figure 2A The method 300 of the process 200. However, the method 300 may additionally or alternatively be performed by any one system or any combination of systems, including but not limited to those described herein.

[0082] Figure 3 300 for generating a trajectory from transformed image data according to some embodiments of the present disclosure. At block B302, the method 300 includes receiving image data generated by a first camera having first camera parameters. For example, the image data 104 may be received after being generated by the first camera of the vehicle 138. The image data 104 may represent a captured image, such as Figure 2A The original image 202 in FIG.

[0083] At block B304, the method 300 includes applying a transformation to the image data to generate transformed image data. For example, a transformation (e.g., a shift, a rotation, etc.) may be applied to the image data to generate transformed image data. The transformation may transform the image data from the perspective of the first camera of the vehicle 138 to the perspective of a reference or ideal camera.

[0084] At block B306, method 300 includes applying the transformed image data to a machine learning model. For example, the transformed image data (e.g., as image data 104) can be applied to one or more machine learning models 108, such as a CNN.

[0085] At block B308, method 300 includes computing output data representing one or more trajectory points by the machine learning model. For example, one or more machine learning models 108 may compute one or more outputs 110 based at least in part on the transformed image data, where the outputs include trajectory data 112. As described herein, trajectory data 112 may represent one or more trajectory points in a 2D or 3D world space along a predicted or recommended trajectory.

[0086] At block B310, the method 300 includes sending the output data to a control component of the vehicle. For example, one or more outputs 110 may be sent to one or more control components 118 of the vehicle 138. In some examples, the one or more control components 118 may use the track points as reference points for navigation. In this way, the one or more control components 118 may determine the steering angle, acceleration, and / or deceleration of the vehicle 138, which may be executed to control the vehicle 138 (or the origin of the vehicle 138) to each track point.

[0087] As the vehicle moves through the environment, method 300 may repeat for each image represented by the image data, as indicated by dashed arrow 312 .

[0088] Although the method 300 is described with respect to image data and a camera, this is not limiting. In some examples, the method 300 can be applied to sensor data generated from one or more sensors in addition to or as an alternative to a camera. For example, sensor data from a LIDAR sensor can be converted to the perspective of a reference or ideal LIDAR sensor. In this way, one or more machine learning models 108 are trained and deployed on sensor data from an ideal or reference sensor.

[0089] Train one or more machine learning models to generate trajectories

[0090] The systems of the present disclosure can use the automatic labeling or annotations as ground truth data to train a machine learning model (e.g., one or more DNNs described herein) to compute track points, vehicle orientation (e.g., relative to features of the environment, such as lane markings), and / or vehicle state (e.g., relative to vehicle maneuvers, such as lane changes, turns, merges, etc.). For example, when collecting image data, additional sensor data that tracks the vehicle's orientation, steering information, and / or speed can also be collected. The additional sensor data can be used to automatically generate (e.g., using a dead reckoning process) tracks—represented by 2D or 3D track points in world space—that can be used as ground truth data for training the DNN. By automatically generating ground truth data using sensor data, the system can be easily expanded by collecting more data—thereby reducing the additional task of manually labeling and annotating any new data that is collected.

[0091] During training data collection, image data, LIDAR data, SONAR data, audio data, and / or other data types may be captured for use as input to the DNN. At the same time, sensor data from steering sensors, IMU sensors, speed sensors, etc. may be captured to reconstruct trajectories (e.g., trajectory points in 2D or 3D world space) corresponding to the image data, LIDAR data, SONAR data, audio data, etc. Thus, during training, the output of the DNN may be compared to the trajectory points in world space - thereby intrinsically training the DNN to account for the conversion from 2D pixel coordinates to 2D or 3D world space coordinates without requiring any pre-processing or post-processing that would require the same.

[0092] In some embodiments, the DNN can be trained to predict state information - such as the state of a vehicle maneuver. In such embodiments, some manual ground truth generation (e.g., manual labeling or annotation) may be required. For example, for lane changes, manual labels associated with images captured during the lane change may be used to indicate progress through the lane change, and for turns, manual labels associated with images captured in the turn may be used to indicate changes through the lane. In such embodiments, the DNN can be trained using the automatically generated trajectory information and the manually labeled state information as ground truth data to learn to compute one or more trajectory points and state information.

[0093] Now refer to Figure 4 , Figure 44 is a data flow diagram illustrating an example process 400 for training one or more machine learning models to generate path planning information according to some embodiments of the present disclosure. In some embodiments, the process 400 can be used to train one or more machine learning models 108 such as described herein. One or more inputs 402 may include training image data 404 and / or training sensor data 406, which may be similar to the image data 104 and / or sensor data 106, respectively, but correspond to training data for training one or more machine learning models 108. As described herein, the training image data 404 and / or training sensor data 406 may be transformed to the perspective of an ideal or reference camera before being input to one or more machine learning models 108. In addition, other types of data pre-processing may be performed herein with respect to the image data 104 and / or sensor data 106, such as but not limited to those described. One or more outputs 408 may include trajectory data 410, vehicle position 412 and / or vehicle state 414, which may be similar to trajectory data 112, vehicle position 114 and / or vehicle state 116, respectively, but may correspond to outputs of one or more machine learning models 108 during training.

[0094] The trajectory training data 418 can be used as ground truth data for training one or more machine learning models 108 (using one or more loss functions 416) to generate one or more outputs 408. When controlling one or more vehicles through the environment, the trajectory training data 418 can be generated by one or more sensors of one or more vehicles that generate training image data 404 and / or training sensor data 406. In some embodiments, the one or more vehicles can be controlled by one or more human drivers through the physical environment, so that the path represented by the trajectory training data 418 represents one or more paths that one or more human drivers have taken. For example, the trajectory training data 418 may include trajectory data generated during an exclusive or almost exclusive human driving portion of the path taken by the vehicle. In some examples, the human driving portion of the path may be from a remotely located human pilot, such as a person driving the vehicle from a remote simulator. Therefore, the trajectory training data 418 may include trajectory data that is automatically labeled and associated with a representation or instance of the sensor data 406 based on associating the perceptual sensor data with the speed, direction, and / or other data generated during the human driving portion of the vehicle operation. Trajectory training data 418 may be used to train one or more machine learning models 108 not only on actual path driving, but also on cues in one or more inputs 402 that may be responsible for producing such paths.

[0095] In other examples, trajectory training data 418 may include annotated trajectories or paths in a 2D image or sensor space and transformed into trajectory labels in 3D world coordinates. For example, a representation or instance (e.g., image, depth map, etc.) of sensor data 406 in coordinates associated with sensor data 406 (e.g., with respect to an image, 2D image coordinates) may be annotated, and the annotations may be transformed into 3D world coordinates (e.g., representing the trajectory through a real-world environment) before being used as ground truth data. However, in some examples, manual labeling may not be performed to account for certain features of the environment (e.g., large objects in adjacent lanes, potholes, etc.). In addition, manual labeling may often result in the middle of lane labels and therefore may not account for these natural maneuvers and changes in paths or trajectories that a human driver may make to account for a particular physical environment. In some examples, a combination of automatically generated trajectories and manual annotations from human driving operations of the vehicle may be used.

[0096] In some embodiments, each vehicle that captures training image data 404 and / or training sensor data 406 (e.g., LIDAR data, SONAR data, etc.) may generate training trajectory data 418 while capturing training trajectory data 418 and / or training sensor data 406. Thus, as the vehicle that captures training image data 404 and / or training sensor data 406 navigates through an environment (e.g., along a road or other driving surface), training trajectory data 418 may be generated by one or more IMU sensors 666, steering sensors from steering system 654, one or more speed sensors 644 (e.g., speedometers), and / or other sensor types to record or reconstruct the path of the vehicle through the environment. For example, training trajectory data 418 may be generated using one or more processes (e.g., dead reckoning processes) to reconstruct a 2D or 3D world space path (e.g., using world space coordinates) of a vehicle through an environment that may correspond at least in part to training image data 404 and / or training sensor data 406. The training trajectory data 418 may represent trajectories or trajectory points that may be used as ground truth data for training one or more machine learning models 108 to generate output 408, such as trajectory data 410, with an acceptable degree of accuracy.

[0097] As a non-limiting example, a camera of a vehicle performing data capture may generate image data representing an image of the physical environment in which the vehicle is located (e.g., an image of the field of view in front of the vehicle). 1 The vehicle may move forward in world space, and during and after capturing the image data, the vehicle may capture training trajectory data 418 using any number of sensors, such as those described herein. Thus, the vehicle may be moving forward in world space, and during and after capturing the image data, the vehicle may capture training trajectory data 418 using any number of sensors, such as those described herein. 1After the time t 2 - Capturing the training trajectory data while the vehicle is traveling through the portion of the physical environment represented in the image. As a result, the training trajectory data 418 can be used to reconstruct the path of the vehicle through the portion of the physical environment in 2D or 3D world space. The path through the portion of the physical environment represented in the image can then be associated with the image (e.g., automatically) as a label or annotation for use as ground truth data.

[0098] For example, the trajectory or trajectory points can be used as labels or annotations corresponding to the training image data 404 and / or the training sensor data 406. In such an example, the reconstructed trajectory 418 from the trajectory training data can be associated with the captured image and / or other data representations from one or more sensors of the vehicle (e.g., LIDAR sensor, RADAR sensor, etc.). The labels or annotations can be automatically generated using the trajectory training data 418, and thus no manual labeling may be required to generate ground truth data. However, in some examples, in addition to automatic labeling or as an alternative, manual labeling can be performed. Therefore, in any example, the label can be synthetically generated (e.g., generated from a computer model or rendering), truly generated (e.g., designed and generated from real-world data), machine automated (e.g., using feature analysis and learning to extract features from the data and then generate labels), human annotation (e.g., a labeler or annotation expert, defining the location of the label) and / or a combination thereof (e.g., the training trajectory data 418 is used to automatically generate labels, and humans correct any errors). In some examples, for each input image, or for each input sensor data representation, there may be a corresponding path label.

[0099] One or more machine learning models 108 may perform forward computations on training image data 404 and / or training sensor data 406. In some examples, one or more machine learning models 108 may learn training trajectories corresponding to 2D or 3D world space coordinates (e.g., represented by trajectory points) corresponding to recommended or predicted trajectories for each image and / or sensor data representation. One or more loss functions 316 may be used to measure the loss (e.g., error) in output trajectory data 410 and / or other outputs 408 (e.g., as predicted by one or more machine learning models 108) compared to ground truth data (e.g., labels or annotations corresponding to trajectory training data 418). For example, a binary cross entropy loss function, a mean squared error (L2) loss function, an L1 loss function, and / or other loss function types may be used as one or more loss functions 316. In some embodiments, two or more different loss functions may be used. For example, one or more loss functions may be used for each type of output 408 where there are two or more outputs 408, or two or more loss functions may be used for a single output type. In the case where two or more loss functions are used for a single output (e.g., for trajectory data 410) type, the loss functions can be weighted relative to each other to generate a weighted loss function. A backward pass calculation can be performed to recursively calculate the gradient of the loss function with respect to the training parameters (e.g., weights, biases, etc.), as shown by arrow 420. In some examples, the weights and biases of one or more machine learning models 108 can be used to calculate these gradients.

[0100] In some examples, as described herein, a distance-based weighted loss can be added to the loss function 316, where the loss function 316 can increasingly penalize losses that are farther from the bottom of the image or other data representation (or from the vehicle). The distance-based weighted loss function penalizes more errors at long distances and thus improves the accuracy of predicted tracks or track points in long distances.

[0101] In some examples, such as where one or more machine learning models 108 predict vehicle orientation 412 and / or vehicle state 414, the training process for these outputs 408 can be similar to U.S. non-provisional application No. 16,241,005, filed on January 7, 2019, the entire contents of which are incorporated herein by reference.

[0102] Thus, one or more machine learning models 108 may be trained to receive image data 104 and / or sensor data 106 as input 102 ( Figure 1A), and generate trajectory data 112, vehicle position 412, vehicle state 414, and / or other outputs 110. In addition, because the one or more machine learning models 408 can be trained using trajectory training data 418 generated during and / or immediately after capturing the one or more inputs 402, the one or more machine learning models 408 can be trained using the trajectory training data 418, and the machine learning models 108 can be trained to generate 2D or 3D world space coordinates as trajectory data 410. In this way, the one or more machine learning models 108 can be trained to implicitly calculate the transformation from 2D image space to 2D or 3D world space using automatically generated labels or annotations from the trajectory training data 418. By implicitly calculating this transformation, the amount of post-processing is reduced, and the output 110 of the one or more machine learning models 108 can be directly used by one or more control components 118, and / or one or more other components of the vehicle 138 (e.g., one or more components of the autonomous driving software stack).

[0103] Reference now Figures 5A-5C , each block of methods 500A, 500B, and 500C described herein includes a computing process that can be performed using any combination of hardware, firmware, and / or software. For example, various functions can be performed by a processor executing instructions stored in a memory. Methods 500A, 500B, and 500C can also be embodied as computer-usable instructions stored on a computer storage medium. Methods 500A, 500B, and 500C can be provided by a stand-alone application, a service or a hosted service (standalone or in combination with another hosted service), or a plug-in to another product, to name a few examples. In addition, as an example, with respect to Figure 4 Methods 500A, 500B, and 500C are described in process 400. However, these methods 500A, 500B, and 500C may additionally or alternatively be performed by any one system or any combination of systems, including but not limited to those described herein.

[0104] Figure 5A 5 is a flow chart illustrating a method 500A for training one or more machine learning models to generate path planning information according to some embodiments of the present disclosure. At block B502, block 500A includes receiving image data representing an image captured at a first time. For example, image data captured by a camera of a vehicle at a time t may be received. 1 Generated training image data.

[0105] At block B504, method 500A includes receiving trajectory training data representing a speed and orientation of a vehicle captured during a period of time including at least a second time after a first time. For example, trajectory training data 418 may be received, wherein trajectory training data 418 is captured during a period of time including at time t 1 At least time t after 2 .

[0106] At block B506, method 500A includes automatically generating a trajectory by determining world space coordinates of the vehicle. For example, trajectory training data 418 may be used to automatically generate a trajectory or trajectory points of the vehicle in a 2D or 3D world space. The trajectory may be associated with the image because the trajectory may represent a path of the vehicle through a portion of the physical environment represented by the image.

[0107] At block B508, method 500A includes training a machine learning model using world space coordinates associated with the image as ground truth data. For example, one or more machine learning models 108 may be trained using world space coordinates generated from trajectory training data 418 as ground truth data (e.g., as labels or annotations associated with the image).

[0108] Figure 5B is a flow chart illustrating a method 500B for training one or more machine learning models to generate path planning information according to some embodiments of the present disclosure. At block B512, the method 500B includes receiving trajectory training data and sensor data representing a sensor field of a sensor. For example, trajectory training data 418 and image data 404 and / or sensor data 406 may be received, where the image data 404 and / or sensor data 406 may represent one or more images, depth maps, SONAR data representations, combinations thereof, and the like.

[0109] At block B514, method 500B includes generating a trajectory for each image of the image. For example, for each image, a trajectory or trajectory points may be generated using trajectory training data. Figure 5C Method 500C may be used to perform block B514. For example, at block B514A, method 500C includes generating a trajectory of the vehicle using a subset of trajectory training data corresponding to a path of the vehicle. For example, a trajectory or trajectory points of the vehicle may be generated using a subset of trajectory training data corresponding to a path of the vehicle through a portion of a physical environment represented in a corresponding sensor data representation. At block B514B, block 500C includes associating the trajectory with a representation of the sensor data. For example, the trajectory or trajectory points may be associated with a corresponding representation of image data 404 (e.g., an image) and / or sensor data 406 as a label or annotation.

[0110] At block B516 , method 500B includes applying the sensor data as input to the machine learning model. For example, the image data 404 and / or the sensor data 406 of the representation may be applied as input 402 to one or more machine learning models 108 .

[0111] At block B518, method 500B includes training the machine learning model based on comparing the output calculated by the machine learning model with the trajectory. For example, the machine learning model 108 can be trained by comparing one or more outputs 408 (e.g., trajectory data 410) of one or more machine learning models 108 with trajectories generated or reconstructed from trajectory training data 418.

[0112] Example Autonomous Vehicle

[0113] Fig. 6A is an illustration of an exemplary autonomous vehicle 138 according to some embodiments of the present disclosure. Autonomous vehicle 138 (optionally referred to herein as "vehicle 138") may include a passenger vehicle, such as a car, truck, bus, and / or other type of vehicle that accommodates one or more passengers. Autonomous vehicles are generally described according to the level of automation, as defined by the National Highway Traffic Safety Administration (NHTSA), (a division of the U.S. Department of Transportation) and the Society of Automotive Engineers (SAE), "Classification and Definitions of Terms Relating to Automated Driving Systems for Highway Motor Vehicles" (Standard No. J3016-201806, issued on June 15, 2018, Standard No. J3016-201609, issued on September 30, 2016, and previous and future versions of the standard). Vehicle 138 may have one or more functions from Level 3 to Level 5 according to the level of automated driving. For example, depending on the embodiment, vehicle 138 may have conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0114] The vehicle 138 may include components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of the vehicle. The vehicle 138 may include a propulsion system 650, such as an internal combustion engine, a hybrid power plant, an all-electric engine, and / or other propulsion system types. The propulsion system 650 may be connected to a drive mechanism of the vehicle 138, which may include a transmission, to achieve propulsion of the vehicle 138. The propulsion system 650 may be controlled in response to receiving a signal from a throttle / accelerator 652.

[0115] When propulsion system 650 is in operation (e.g., when the vehicle is in motion), vehicle 138 may be guided (e.g., along a desired path or route) using steering system 654, which may include a steering wheel. Steering system 654 may receive signals from steering actuator 656. A steering wheel may be optional for fully automated (Level 5) functionality.

[0116] Brake sensor system 646 may be used to operate vehicle brakes in response to receiving signals from brake actuator 648 and / or brake sensors.

[0117] The one or more controllers 636 may include one or more system on chip (SoC) 604 ( Figure 6C ) and / or GPU, can provide signals (e.g., command representations) to one or more components. For example, one or more controllers can send signals via one or more brake actuators 648 to operate vehicle brakes, via one or more steering actuators 656 to operate steering system 654, and via one or more throttle / accelerators 652 to operate propulsion system 650. One or more controllers 636 may include one or more onboard (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals, which process sensor data and output operating commands (e.g., signals representing commands) to achieve autonomous driving and / or assist a human driver in driving the vehicle 138. One or more controllers 636 may include a first controller 636 for autonomous driving functions, a second controller 636 for functional safety functions, a third controller 636 for artificial intelligence functions (e.g., computer vision), a fourth controller 636 for infotainment program functions, a fifth controller 636 for redundancy in emergency situations, and / or other controllers. In some examples, a single controller 636 may handle two or more of the above functions, two or more controllers 636 may handle a single function, and / or any combination thereof.

[0118] The one or more controllers 636 may provide signals for controlling one or more components and / or systems of the vehicle 138 in response to sensor data received from one or more sensors (eg, sensor inputs). Sensor data may be received from, for example, but not limited to, one or more global navigation satellite system sensors 658 (e.g., one or more global positioning system sensors), one or more RADAR sensors 660, one or more ultrasonic sensors 662, one or more LIDAR sensors 664, one or more inertial measurement unit (IMU) sensors 666 (e.g., one or more accelerometers, one or more gyroscopes, one or more magnetic compasses, one or more magnetometers, etc.), one or more microphones 696, one or more stereo cameras 668, one or more wide field of view cameras 670 (e.g., fisheye cameras), one or more infrared cameras 672, one or more surround cameras 674 (e.g., 360 degree cameras), one or more long-range and / or mid-range cameras 698, one or more speed sensors 644 (e.g., for measuring the speed of the vehicle 138), one or more vibration sensors 642, one or more steering sensors 640, one or more brake sensors (e.g., as part of a brake sensor system 646), and / or other sensor types.

[0119] The one or more controllers 636 may receive input (e.g., represented by input data) and provide output (e.g., represented by output data, display data, etc.) from the instrument cluster 632 of the vehicle 138 via a human machine interface (HMI) display 634, an audible annunciator, a speaker, and / or via other components of the vehicle 138. The output may include information such as vehicle speed, velocity, time, map data (e.g., Figure 6C The HMI display 634 may include information about the presence of one or more objects (e.g., street signs, warning signs, traffic light changes, etc.), and / or information about maneuvers that have been made, are being made, or are about to be made (e.g., changing lanes now, taking exit 34B in two miles, etc.).

[0120] The vehicle 138 also includes a network interface 624 that can use one or more wireless antennas 626 and / or one or more modems to communicate through one or more networks. For example, the network interface 624 can communicate through LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The one or more wireless antennas 626 can also use one or more local area networks, such as Bluetooth, Bluetooth LE, Z-Wave, ZigBee, etc., and / or one or more low power wide area networks (LPWANs), such as LoRaWAN, SigFox, etc. to achieve communication between objects in the environment (e.g., vehicles, mobile devices, etc.).

[0121] Figure 6B According to some embodiments of the present disclosure, Fig. 6A 1. An example of camera positions and fields of view for an exemplary autonomous vehicle 138. The cameras and corresponding fields of view are an example embodiment and are not intended to be limiting. For example, additional and / or alternative cameras may be included, and / or the cameras may be located at different locations on the vehicle 138.

[0122] The camera type used for the camera may include, but is not limited to, a digital camera, which may be applicable to components and / or systems of the vehicle 138. One or more cameras may operate at automotive safety integrity level (ASIL) B and / or another ASIL. Depending on the embodiment, the camera type may have any image capture rate, such as 60 frames per second (fps), 620fps, 240fps, etc. The camera may be able to use a rolling shutter, a global shutter, other types of shutters, or a combination thereof. In some examples, the color filter array may include a red transparent light (RCCC) filter array, a red transparent light blue (RCCB) filter array, a red blue green transparent (RBGC) filter array and a Foveon X3 color filter array, a Bayer sensor (RGGB) filter array, a monochrome sensor filter array, and / or other types of filter arrays. In some embodiments, a transparent pixel camera, such as a camera with an RCCC, RCCB, and / or RBGC filter array, may be used in an effort to increase light sensitivity.

[0123] In some examples, one or more cameras may be used to perform advanced driver assistance system (ADAS) functions (e.g., as part of a redundant or fail-safe design). For example, a multi-function mono camera may be installed to provide functions including lane departure warning, traffic sign assistance, and intelligent headlight control. One or more cameras (e.g., all cameras) may record and provide image data (e.g., video) simultaneously.

[0124] One or more cameras can be mounted in a mounting assembly, such as a custom designed (3-D printed) assembly, to remove stray light and reflections in the vehicle (e.g., reflections from the dashboard reflected in the windshield mirror) that may interfere with the camera's image data capture function. Referring to the wing mounting assembly, the wing mounting assembly can be custom made for 3D printing so that the camera mounting plate matches the shape of the wing mirror. In some examples, one or more cameras can be integrated into the wing profile. For side-view cameras, one or more cameras can also be integrated into the four pillars at each corner of the cabin.

[0125] A camera with a field of view that includes a portion of the environment in front of the vehicle 138 (e.g., a front-facing camera) can be used for a surround view to help identify oncoming paths and obstacles ahead, as well as, with the help of one or more controllers 636 and / or control SoCs, assist in providing information critical to generating an occupancy grid and / or determining a preferred vehicle path. A front-facing camera can be used to perform many of the same ADAS functions as LIDAR, including emergency braking, pedestrian detection, and collision avoidance. A front-facing camera can also be used for ADAS functions and systems, including lane departure warning (“LDW”), adaptive cruise control (“ACC”), and / or other functions, such as traffic sign recognition.

[0126] A variety of cameras may be used in a front-facing configuration, including, for example, a monocular camera platform including a CMOS (complementary metal oxide semiconductor) color imager. Another example may be one or more wide field of view cameras 670, which may be used to sense objects entering the field of view from the periphery (e.g., pedestrians, crossing traffic, or bicycles). Although Figure 6B Only one wide-view camera is shown in FIG. 1 , and any number of wide-view cameras 670 may be present on the vehicle 138. In addition, one or more remote cameras 698 (e.g., a long-view stereo camera pair) may be used for depth-based object detection, particularly for objects for which a neural network has not yet been trained. One or more remote cameras 698 may also be used for object detection and classification, as well as basic object tracking.

[0127] One or more stereo cameras 668 may also be included in the front configuration. One or more stereo cameras 668 may include an integrated control unit including a scalable processing unit that may provide programmable logic (FPGA) and a multi-core microprocessor with an integrated CAN or Ethernet interface on a single chip. Such a unit may be used to generate a 3-D map of the vehicle environment, including distance estimates for all points in the image. Alternative one or more stereo cameras 668 may include one or more compact stereo vision sensors, which may include two camera lenses (one each on the left and right) and an image processing chip that may measure the distance from the vehicle to the target object, and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. In addition to those described herein or as an alternative, one or more stereo cameras 668 of other types may be used.

[0128] The camera's field of view includes a portion of the environment to the side of the vehicle 138 (e.g., a side view camera), which can be used for surround view, providing information for creating and updating occupancy grids, and for generating side collision warnings. For example, one or more surround cameras 674 (e.g., such as Figure 6B The four surround cameras 674 shown in FIG. 1 may be positioned on the vehicle 138. The one or more surround cameras 674 may include one or more wide field of view cameras 670, one or more fisheye cameras, one or more 360 ​​degree cameras, and / or the like. For example, four fisheye cameras may be positioned on the front, rear, and sides of the vehicle. In an alternative arrangement, the vehicle may use three surround cameras 674 (e.g., left, right, and rear), and may utilize one or more other cameras (e.g., a front-facing camera) as a fourth surround view camera.

[0129] A camera having a field of view that includes a portion of the environment to the rear of the vehicle 138 (e.g., a rear view camera) can be used for parking assistance, surround view, rear collision warning, and creating and updating an occupancy grid. As described herein, a variety of cameras can be used, including but not limited to cameras that are also suitable as front-facing cameras (e.g., one or more long-range and / or mid-range cameras 698, one or more stereo cameras 668, one or more infrared cameras 672, etc.).

[0130] Figure 6C According to some embodiments of the present disclosure Fig. 6ABlock diagram of an example system architecture for an exemplary autonomous vehicle 138. It should be understood that this and other arrangements described herein are set forth only as examples. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, functional groupings, etc.) may be used in addition to or in place of those shown, and some elements may be omitted entirely. In addition, many of the elements described herein are functional entities that may be implemented as discrete or distributed components or in conjunction with other components, and may be implemented in any suitable combination and location. The various functions described herein as being performed by the entities may be performed by hardware, firmware, and / or software. For example, the various functions may be performed by a processor executing instructions stored in a memory.

[0131] Figure 6C Each component, feature, and system in the vehicle 138 is shown connected via a bus 602. The bus 602 may include a controller area network (CAN) data interface (optionally referred to herein as a "CAN bus"). The CAN can be a network within the vehicle 138 that assists in controlling various features and functions of the vehicle 138, such as actuation of brakes, acceleration, braking, steering, windshield wipers, etc. The CAN bus can be configured to have dozens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus can be read to find steering wheel angle, ground speed, engine revolutions per minute (RPM), button position, and / or other vehicle status indicators. The CAN bus can comply with ASIL B.

[0132] Although bus 602 is described as a CAN bus here, this is not restrictive. For example, FlexRay and / or Ethernet can be used in addition to or as an alternative to the CAN bus. In addition, although a single line is used to represent bus 602, this is not restrictive. For example, there can be any number of buses 602, which can include one or more CAN buses, one or more FlexRay buses, one or more Ethernet buses, and / or one or more other types of buses using different protocols. In some examples, two or more buses 602 can be used to perform different functions, and / or can be used for redundancy. For example, a first bus 602 can be used for collision avoidance functions, and a second bus 602 can be used for actuation control. In any example, each bus 602 can communicate with any component of the vehicle 138, and two or more buses 602 can communicate with the same component. In some examples, each SoC 604, each controller 636, and / or each computer in the vehicle can access the same input data (e.g., input from sensors of the vehicle 138), and can be connected to a common bus, such as a CAN bus.

[0133] The vehicle 138 may include one or more controllers 636, such as those described herein. Fig. 6A The controllers described above may be used for various functions. The controller(s) 636 may be coupled to any of the various other components and systems of the vehicle 138 and may be used to control the vehicle 138, the artificial intelligence of the vehicle 138, the entertainment information programs of the vehicle 138, etc.

[0134] The vehicle 138 may include one or more systems on a chip (SoC) 604. The SoC 604 may include one or more CPUs 606, one or more GPUs 608, one or more processors 610, one or more caches 612, one or more accelerators 614, one or more data stores 616, and / or other components and features not shown. The one or more SoCs 604 may be used to control the vehicle 138 in a variety of platforms and systems. For example, the one or more SoCs 604 may be combined with an HD map 622 in a system (e.g., a system of the vehicle 138), and the HD map 622 may be downloaded from one or more servers (e.g., Fig.6D server 678) to obtain map refreshes and / or updates.

[0135] One or more CPUs 606 may include a CPU cluster or CPU complex (optionally referred to herein as a "CCPLEX"). One or more CPUs 606 may include multiple cores and / or L2 caches. For example, in some embodiments, one or more CPUs 606 may include eight cores in a multiprocessor configuration coupled to each other. In some embodiments, one or more CPUs 606 may include four dual-core clusters, each with a dedicated L2 cache (e.g., a 2MB L2 cache). One or more CPUs 606 (e.g., CCPLEX) may be configured to support simultaneous cluster operations so that any combination of CPU 606 clusters is active at any given time.

[0136] One or more CPUs 606 may implement power management functional capabilities including one or more of the following features: individual hardware blocks may be automatically clock gated when idle to save dynamic power; each core clock may be gated when the core is not actively executing instructions due to executing WFI / WFE instructions; each core may be independently power gated; each core cluster may be independently clock gated when all cores are clock gated or power gated; and / or each core cluster may be independently power gated when all cores are power gated. One or more CPUs 606 may also implement an enhanced algorithm for managing power states, where allowed power states and expected wakeup times are specified, and hardware / microcode determines the optimal power state for core, cluster, and CCPLEX inputs. The processing core may support a simplified power state entry sequence in software, where work is offloaded to microcode.

[0137] One or more GPUs 608 may include an integrated GPU (optionally referred to herein as an "iGPU"). One or more GPUs 608 may be programmable and may be effective for parallel workloads. In some examples, one or more GPUs 6 08 may use an enhanced tensor instruction set. One or more GPUs 608 may include one or more streaming microprocessors, each of which may include an L1 cache (e.g., an L1 cache with at least 96KB storage capacity), and two or more streaming microprocessors may share an L2 cache (e.g., an L2 cache with 512KB storage capacity). In some embodiments, one or more GPUs 608 may include at least eight streaming microprocessors. One or more GPUs 608 may use one or more computing application programming interfaces (APIs). In addition, one or more GPUs 608 may use one or more parallel computing platforms and / or programming models (e.g., NVIDIA's CUDA).

[0138] One or more GPUs 608 may be power optimized to achieve optimal performance in automotive and embedded use cases. For example, one or more GPUs 608 may be manufactured on Fin field effect transistors (FinFETs). However, this is not restrictive, and one or more GPUs 608 may be manufactured using other semiconductor manufacturing processes. Each streaming microprocessor may include multiple mixed precision processing cores divided into multiple blocks. For example, but not limited to, 64 PF32 cores and 32 PF64 cores may be divided into four processing blocks. In such an example, 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed precision NVIDIA TENSOR COREs for deep learning matrix arithmetic, L0 instruction cache, thread warp scheduler, dispatch unit, and / or 64KB register files may be allocated to each processing block. In addition, the streaming microprocessor may include independent parallel integer and floating point data paths to provide efficient execution of workloads through mixed calculations and addressing calculations. The streaming microprocessor may include independent thread scheduling capabilities to achieve finer-grained synchronization and collaboration between parallel threads. A streaming microprocessor may include a combined L1 data cache and shared memory unit to increase performance while simplifying programming.

[0139] In some examples, one or more GPUs 608 may include high bandwidth memory (HBM) and / or a 16GB HBM2 memory subsystem to provide a peak memory bandwidth of approximately 900 GB / sec. In some examples, synchronous graphics random access memory (SGRAM), such as graphics double data rate type five synchronous random access memory (GDDR5), may be used in addition to or in lieu of HBM memory.

[0140] One or more GPUs 608 may include unified memory technology that includes access counters to allow more accurate migration of memory pages to the processors that access them most frequently, thereby improving the efficiency of memory ranges shared between processors. In some examples, address translation service (ATS) support may be used to allow one or more GPUs 608 to directly access one or more CPU 606 page tables. In such an example, when one or more GPU 608 memory management units (MMUs) experience a miss, an address translation request may be sent to one or more CPUs 606. In response, one or more CPUs 606 may look up a table for virtual-to-physical mapping of addresses in its pages and send the translation back to one or more GPUs 608. Therefore, unified memory technology may allow a single unified virtual address space to be used for memory of one or more CPUs 606 and one or more GPUs 608, thereby simplifying application programming and porting of one or more GPUs 608 to one or more GPUs 608.

[0141] Additionally, one or more GPUs 608 may include access counters that can track the frequency of accesses to other processors' memory by one or more GPUs 608. The access counters can help ensure that memory pages are moved to the physical memory of the processor that accesses the page most frequently.

[0142] One or more SoCs 604 may include any number of one or more caches 612, including those described herein. For example, one or more caches 612 may include an L3 cache (e.g., a cache that connects both one or more CPUs 606 and one or more GPUs 608) that can be used for one or more CPUs 606 and one or more GPUs 608. One or more caches 612 may include a write-back cache that can track the state of a line, such as by using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). Depending on the embodiment, the L3 cache may include 4MB or more, although smaller cache sizes may be used.

[0143] One or more SoCs 604 may include one or more accelerators 614 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, one or more SoCs 604 may include a hardware acceleration cluster, which may include optimized hardware accelerators and / or large on-chip memory. Large on-chip memory (e.g., 4MB of SRAM) can enable hardware acceleration clusters to accelerate neural networks and other calculations. Hardware acceleration clusters can be used to supplement one or more GPUs 608 and unload some tasks of one or more GPUs 608 (e.g., to free up more cycles of one or more GPUs 608 to perform other tasks). As an example, accelerators 614 can be used for target workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are stable enough to be suitable for acceleration. The term "CNN" used herein may include all types of CNNs, including region-based or region-based convolutional neural networks (RCNNs) and fast RCNNs (e.g., for object detection).

[0144] One or more accelerators 614 (e.g., a hardware acceleration cluster) may include one or more deep learning accelerators (DLAs). One or more DLAs may include one or more tensor processing units (TPUs) that may be configured to provide an additional 10 trillion operations per second for deep learning applications and reasoning. A TPU may be an accelerator configured and optimized to perform image processing functions (e.g., for CNN, RCNN, etc.). One or more DLAs may also be optimized for a specific set of neural network types and floating point operations and reasoning. The design of one or more DLAs may provide more performance per millimeter than a general-purpose GPU and far exceed the performance of a CPU. One or more TPUs may perform several functions, including single instance convolution functions, such as supporting INT8, INT16, and FP16 data types for features and weights and post-processor functions.

[0145] One or more DLAs can quickly and efficiently execute neural networks, particularly CNNs, on processed or unprocessed data for any of a variety of functions, including, for example, but not limited to: a CNN for object identification and detection using data from a camera sensor; a CNN for distance estimation using data from a camera sensor; a CNN for emergency vehicle detection and identification and detection using data from a microphone; a CNN for facial recognition and vehicle owner recognition using data from a camera sensor; and / or a CNN for safety and / or security-related events.

[0146] One or more DLAs can perform any function of one or more GPUs 608, and by using inference accelerators, for example, designers can target any function to one or more DLAs or one or more GPUs 608. For example, designers can focus processing of CNNs and floating point operations on one or more DLAs and leave other functions to one or more GPUs 608 and / or other accelerators 614.

[0147] One or more accelerators 614 (e.g., a hardware acceleration cluster) may include one or more programmable vision accelerators (PVAs), which may be referred to herein alternatively as computer vision accelerators. One or more PVAs may be designed and configured to accelerate computer vision algorithms for advanced driver assistance systems (ADAS), autonomous driving, and / or augmented reality (AR) and / or virtual reality (VR) applications. One or more PVAs may provide a balance between performance and flexibility. For example, each PVA may include, for example but not limited to, any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors.

[0148] The RISC core can interact with an image sensor (e.g., an image sensor of any camera described herein), one or more image signal processors, etc. Each RISC core may include any number of memories. Depending on the embodiment, the RISC core may use any of a number of protocols. In some examples, the RISC core may execute a real-time operating system (RTOS). The RISC core may be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC core may include an instruction cache and / or a tightly coupled RAM.

[0149] The DMA may enable components of one or more PVAs to access system memory independently of the one or more CPUs 606. The DMA may support any number of features for providing optimizations to the PVA, including, but not limited to, support for multi-dimensional addressing and / or circular addressing. In some examples, the DMA may support up to six or more dimensions of addressing, which may include block width, block height, block depth, horizontal block stepping, vertical block stepping, and / or depth stepping.

[0150] A vector processor may be a programmable processor that may be designed to efficiently and flexibly execute programming for computer vision algorithms and provide signal processing capabilities. In some examples, the PVA may include a PVA core and two vector processing subsystem partitions. The PVA core may include a processor subsystem, one or more DMA engines (e.g., two DMA engines), and / or other peripherals. The vector processing subsystem may operate as the main processing engine of the PVA and may include a vector processing unit (VPU), an instruction cache, and / or a vector memory (e.g., VMEM). The VPU core may include a digital signal processor, such as a single instruction, multiple data (SIMD), a very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW may increase throughput and speed.

[0151] Each vector processor can include an instruction cache and can be coupled to a dedicated memory. As a result, in some examples, each vector processor can be configured to execute independently of the other vector processors. In other examples, the vector processors included in a particular PVA can be configured to exploit data parallelism. For example, in some embodiments, multiple vector processors included in a single PVA can execute the same computer vision algorithm, but on different areas of an image. In other examples, the vector processors included in a particular PVA can execute different computer vision algorithms simultaneously on the same image, or even execute different algorithms on sequential images or portions of images. Among them, any number of PVAs can be included in the hardware acceleration cluster, and any number of vector processors can be included in each PVA. In addition, one or more PVAs can include additional error correction code (ECC) memory to enhance overall system security.

[0152] One or more accelerators 614 (e.g., a hardware acceleration cluster) may include an on-chip computer vision network and SRAM to provide high-bandwidth, low-latency SRAM to one or more accelerators 614. In some examples, the on-chip memory may include at least 4MB of SRAM, including, for example, but not limited to, eight field-configurable memory blocks that are accessible to both the PVA and DLA. Each pair of memory blocks may include an advanced peripheral bus (APB) interface, configuration circuits, controllers, and multiplexers. Any type of memory may be used. The PVA and DLA may access the memory through a backbone network that provides high-speed access to the memory for the PVA and DLA. The backbone network may include an on-chip computer vision network that interconnects the PVA and DLA to the memory (e.g., using APB).

[0153] The on-chip computer vision network may include an interface that determines that both the PVA and DLA provide ready and valid signals before transmitting any control signals / addresses / data. Such an interface may provide separate phases and separate channels for sending control signals / addresses / data, as well as burst type communications for continuous data transmission. This type of interface may conform to ISO26262 or IEC 61508 standards, although other standards and protocols may be used.

[0154] In some examples, one or more SoCs 604 may include a real-time ray tracing hardware accelerator, such as described in U.S. Patent Application No. 16101232 filed on August 10, 2018. The real-time ray tracing hardware accelerator may be used to quickly and efficiently determine the position and range of objects (e.g., within a world model) to generate real-time visualization simulations, for RADAR signal interpretation, for sound propagation synthesis and / or analysis, for SONAR system simulations, for general wave propagation simulations, for comparison with LIDAR data for positioning and / or other functions, and / or for other purposes.

[0155] One or more accelerators 614 (e.g., a cluster of hardware accelerators) have a wide range of uses for autonomous driving. PVA can be a programmable vision accelerator that can be used in key processing stages in ADAS and autonomous vehicles. The functionality of PVA is well suited for algorithmic domains that require predictable processing at low power and low latency. In other words, PVA performs well on semi-intensive or intensive conventional computations, even on small data sets, which require predictable runtimes with low latency and low power. Therefore, in the context of a platform for autonomous vehicles, PVA is designed to run classic computer vision algorithms because they are efficient at object detection and integer math operations.

[0156] For example, according to one embodiment of the present technology, PVA is used to perform computer stereo vision. In some examples, algorithms based on semi-global matching can be used, but this is not limiting. Many applications for level 3-5 autonomous driving require dynamic estimation / stereo matching on the fly (e.g., structure from motion, pedestrian recognition, lane detection, etc.). PVA can perform computer stereo vision functions on input from two monocular cameras.

[0157] In some examples, PVA can be used to perform dense optical flow. Processed RADAR is provided based on processing raw RADAR data (e.g., using a 4D fast Fourier transform). In other examples, PVA is used for time-of-flight depth processing, for example, by processing raw time-of-flight data to provide processed time-of-flight data.

[0158] DLA can be used to run any type of network to enhance control and driving safety, including, for example, a neural network that outputs a confidence measure for each object detection. Such a confidence value can be interpreted as a probability, or to provide a relative "weight" of each detection compared to other detections. This confidence value enables the system to make further decisions about which detections should be considered true positive detections rather than false positive detections. For example, the system can set a threshold for confidence and only consider detections that exceed the threshold as true positive detections. In an automatic emergency braking (AEB) system, a false positive detection will cause the vehicle to automatically perform emergency braking, which is obviously undesirable. Therefore, only the most confident detections should be considered as triggers for AEB. DLA can run a neural network to regress confidence values. The neural network may take as its input at least some subset of parameters, such as bounding box dimensions, a ground plane estimate obtained (e.g., from another subsystem), an inertial measurement unit (IMU) sensor 666 output related to the orientation of the vehicle 138, range, 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., a LIDAR sensor 664 or a RADAR sensor 660), and the like.

[0159] One or more SoCs 604 may include one or more data stores 616 (e.g., memory). One or more data stores 616 may be on-chip memory of the SoC 604 that may store a neural network to be executed on the GPU and / or DLA. In some examples, one or more data stores 616 may be large enough to store multiple instances of a neural network for redundancy and safety. One or more data stores 612 may include one or more L2 or L3 caches 612. References to one or more data stores 616 may include references to memory associated with the PVA, DLA, and / or other accelerators 614, as described herein.

[0160] One or more SoCs 604 may include one or more processors 610 (e.g., embedded processors). One or more processors 610 may include a boot and power management processor, which may be a dedicated processor and subsystem to handle boot power and management functions and related safety implementations. The boot and power management processor may be part of the SoC 604 boot sequence and may provide runtime power management services. The boot power and management processor may provide clock and voltage programming, assist in system low power state transitions, management of one or more SoC 604 thermal and temperature sensors, and / or management of SoC 604 power states. Each temperature sensor may be implemented as a ring oscillator whose output frequency is proportional to the temperature, and one or more SoCs 604 may use the ring oscillator to detect the temperature of one or more CPUs 606, one or more GPUs 608, and one or more accelerators. If it is determined that the temperature exceeds a threshold, the boot and power management processor may enter a temperature fault routine and place one or more SoCs 604 in a lower power state and / or place the vehicle 138 in a driver to safe stop mode (e.g., stopping the vehicle 138 safely).

[0161] The one or more processors 610 may also include a set of embedded processors that may be used as an audio processing engine. The audio processing engine may be an audio subsystem that implements full hardware support for multi-channel audio through multiple interfaces, as well as a wide and flexible range of audio I / O interfaces. In some examples, the audio processing engine is a dedicated processor core having a digital signal processor with dedicated RAM.

[0162] One or more processors 610 may also include an always-on processor engine that may provide the necessary hardware features to support low-power sensor management and wake-up use cases. The always-on processor engine may include a processor core, tightly coupled RAM, supporting peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0163] One or more processors 610 may also include a safety cluster engine that includes a dedicated processor subsystem to handle safety management for automotive applications. The safety cluster engine may include two or more processor cores, tightly coupled RAM, supporting peripherals (e.g., timers, interrupt controllers, etc.), and / or routing logic. In safety mode, two or more cores may operate in lockstep mode and act as a single core with comparison logic to detect any differences between their operations.

[0164] The one or more processors 610 may also include a real-time camera engine, which may include a dedicated processor subsystem for handling real-time camera management.

[0165] The one or more processors 610 may also include a high dynamic range signal processor, which may include an image signal processor, which is a hardware engine that is part of the camera processing transmission path.

[0166] One or more processors 610 may include a video image synthesizer, which may be a processing block (e.g., implemented on a microprocessor) that implements the video post-processing functions required by the video playback application to produce the final image of the player window. The video image synthesizer may perform lens distortion correction on one or more wide field of view cameras 670, one or more surround cameras 674, and / or in-cabin surveillance camera sensors. The in-cabin surveillance camera sensors are preferably monitored by a neural network running on another instance of an advanced SoC, configured to recognize and respond accordingly to cabin events. The in-cabin system may perform lip reading to activate cellular service and place a call, dictate an email, change the vehicle's destination, activate or change the vehicle's infotainment system and settings, or provide voice-activated web surfing. Certain functions are available to the driver only when the vehicle is operating in autonomous mode and are disabled otherwise.

[0167] The video image synthesizer may include enhanced temporal noise reduction for both spatial and temporal noise reduction. For example, in the case of motion in the video, the noise reduction appropriately weights the spatial information, thereby reducing the weight of information provided by adjacent frames. In the case where an image or portion of an image does not include motion, the temporal noise reduction performed by the video image synthesizer may use information from previous images to reduce noise in the current image.

[0168] The video image compositor may also be configured to perform stereo correction on the input stereoscopic footage frames. The video image compositor may also be used for user interface composition when the operating system desktop is in use, and the one or more GPUs 608 do not need to continuously render new surfaces. Even when the one or more GPUs 608 are powered on and actively performing 3D rendering, the video image compositor may be used to offload the one or more GPUs 608 to improve performance and responsiveness.

[0169] One or more SoC 604 may also include a mobile industry processor interface (MIPI) camera serial interface for receiving video and input from a camera, a high-speed interface and / or a video input block that may be used for a camera and associated pixel input functionality. One or more SoC 604 may also include one or more input / output controllers that may be controlled by software and may be used to receive I / O signals that are not committed to a specific role.

[0170] One or more SoCs 604 may also include a wide range of peripheral interfaces to enable communication with peripherals, audio codecs, power management, and / or other devices. One or more SoCs 604 may be used to process data from cameras (e.g., connected via Gigabit multimedia serial links and Ethernet), sensors (e.g., one or more LIDAR sensors 664, one or more RADAR sensors 660, etc., which may be connected via Ethernet), data from bus 602 (e.g., speed of vehicle 138, steering wheel position, etc.), data from one or more GNSS sensors 658 (e.g., connected via Ethernet or CAN bus). One or more SoCs 604 may also include dedicated high-performance mass storage controllers, which may include their own DMA engines and may be used to free up the CPU 606 from routine data management tasks.

[0171] One or more SoCs 604 can be an end-to-end platform with a flexible architecture spanning automation levels 3-5, providing a comprehensive functional safety architecture that, along with deep learning tools, leverages and effectively uses computer vision and ADAS technologies for diversity and redundancy, providing a platform for a flexible, reliable driving software stack. One or more SoCs 604 can be faster, more reliable, and even more energy-efficient and space-efficient than traditional systems. For example, one or more accelerators 614, when combined with one or more CPUs 606, one or more GPUs 608, and one or more data stores 616, can provide a fast, effective platform for level 3-5 autonomous vehicles.

[0172] As a result, the technology provides capabilities and functionality that are not possible with conventional systems. For example, computer vision algorithms can be executed on multiple CPUs, which can be configured using high-level programming languages ​​such as the C programming language to perform various processing algorithms across a variety of visual data. However, CPUs generally cannot meet the performance requirements of many computer vision applications, such as those related to execution time and power consumption. In particular, many CPUs are unable to execute complex object detection algorithms in real time, which is a requirement for in-vehicle ADAS applications and a requirement for practical Level 3-5 autonomous vehicles.

[0173] Compared to conventional systems, by providing a CPU complex, a GPU complex, and a hardware acceleration cluster, the techniques described herein allow multiple neural networks to be executed simultaneously and / or sequentially, and the results to be combined together to enable Level 3-5 autonomous driving capabilities. For example, a CNN executed on a DLA or dGPU (e.g., one or more GPUs 620) can include text and word recognition, allowing the supercomputer to read and understand traffic signs, including signs for which the neural network has not been specifically trained. The DLA can also include a neural network that is able to recognize, interpret, and provide semantic understanding of the symbols, and pass that semantic understanding to a path planning module running on the CPU complex.

[0174] As another example, multiple neural networks may be run simultaneously, as may be required for Level 3, 4, or 5 autonomous driving. For example, a warning sign consisting of “CAUTION: FLASHING LIGHTS INDICATE ICE CONDITIONS” along with electric lights may be interpreted by several neural networks, either independently or collectively. The sign itself may be recognized as a traffic sign by a first deployed neural network (e.g., an already trained neural network), the text “FLASHING LIGHTS INDICATE ICE CONDITIONS” may be interpreted by a second deployed neural network, which notifies the vehicle’s path planning software (preferably executing on a CPU complex) that an icy condition exists when flashing lights are detected. Flashing lights may be identified by operating a third deployed neural network over multiple frames, thereby notifying the vehicle’s path planning software of the presence (or absence) of flashing lights. All three neural networks may be run simultaneously, such as within a DLA and / or on one or more GPUs 608.

[0175] In some examples, a CNN for facial recognition and owner identification can use data from the camera sensor to identify the presence of an authorized driver and / or owner of the vehicle 138. The always-on sensor processing engine can be used to unlock the vehicle when the owner approaches the driver's door and turns on the lights, and in security mode, disable the vehicle when the owner leaves the vehicle. In this way, one or more SoCs 604 provide security against theft and / or carjacking.

[0176] In another example, a CNN for emergency vehicle detection and identification can use data from microphone 696 to detect and identify emergency vehicle sirens. Compared to conventional systems that use general classifiers to detect sirens and manually extract features, one or more SoCs 604 use CNNs to classify environmental and urban sounds, as well as classify visual data. In a preferred embodiment, the CNN running on the DLA is trained to identify the relative closing speed of emergency vehicles (e.g., by using the Doppler effect). The CNN can also be trained to identify emergency vehicles specific to the local area in which the vehicle is operating, as identified by one or more GNSS sensors 658. Thus, for example, when operating in Europe, the CNN will seek to detect European sirens, while in the United States, the CNN will seek to identify only North American sirens. Once an emergency vehicle is detected, a control program can be used to execute an emergency vehicle safety routine to slow the vehicle, pull to the side of the vehicle, park the vehicle, and / or idle the vehicle with the assistance of ultrasonic sensors 662 until one or more emergency vehicles pass.

[0177] The vehicle may include one or more CPUs 618 (e.g., discrete one or more CPUs or one or more dCPUs) that may be coupled to one or more SoCs 604 via a high-speed interconnect (e.g., PCIe). For example, the one or more CPUs 618 may include an X86 processor. The one or more CPUs 618 may be used to perform any of a variety of functions, including, for example, arbitrating potentially inconsistent results between multiple ADAS sensors and one or more SoCs 604, and / or monitoring the status and health of one or more controllers 636 and / or the infotainment system SoC 630.

[0178] The vehicle 138 may include one or more GPUs 620 (e.g., one or more discrete GPUs or one or more dGPUs) that may be coupled to the SoC 604 via a high-speed interconnect (e.g., NVIDIA's NVLINK). The one or more GPUs 620 may provide additional artificial intelligence functionality, such as by executing redundant and / or different neural networks, and may be used to train and / or update the neural network based on input from sensors of the vehicle 138 (e.g., sensor data).

[0179] The vehicle 138 may also include a network interface 624, which may include one or more wireless antennas 626 (e.g., one or more wireless antennas for different communication protocols, such as a cellular antenna, a Bluetooth antenna, etc.). The network interface 624 may be used to wirelessly connect to the cloud (e.g., to one or more servers 678 and / or other network devices), to other vehicles, and / or to computing devices (e.g., a passenger's client device) via the Internet. To communicate with other vehicles, a direct link may be established between the two vehicles and / or an indirect link may be established (e.g., across a network and via the Internet). A vehicle-to-vehicle communication link may be used to provide a direct link. The vehicle-to-vehicle communication link may provide the vehicle 138 with information about vehicles in the vicinity of the vehicle 138 (e.g., vehicles in front of, to the side of, and / or behind the vehicle 138). This functionality may be part of a cooperative adaptive cruise control function for the vehicle 138.

[0180] The network interface 624 may include a SoC that provides modulation and demodulation functions and enables one or more controllers 636 to communicate over a wireless network. The network interface 624 may include a radio frequency front end for up-conversion from baseband to radio frequency, and down-conversion from radio frequency to baseband. The frequency conversion may be performed by well-known processes, and / or may be performed using a superheterodyne process. In some examples, the radio frequency front end function may be provided by a separate chip. The network interface may include wireless functions for communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0181] One or more vehicles 138 may also include one or more data stores 628, which may include off-chip (e.g., off-chip one or more SoCs 604) storage. The one or more data stores 628 may include one or more storage elements, including RAM, SRAM, DRAM, VRAM, flash memory, a hard disk, and / or other components and / or devices that can store at least one bit of data.

[0182] The vehicle 138 may also include one or more GNSS sensors 658. One or more GNSS sensors 658 (e.g., GPS and / or multiple assisted GPS sensors) to assist with mapping, perception, occupancy grid generation, and / or path planning functions. Any number of one or more GNSS sensors 658 may be used, including, for example but not limited to, GPS using a USB connector with an Ethernet to serial (RS-232) bridge.

[0183] The vehicle 138 may also include one or more RADAR sensors 660. The vehicle 138 may use one or more RADAR sensors 660 for remote vehicle detection even in darkness and / or inclement weather conditions. The RADAR functional safety level may be ASIL B. One or more RADAR sensors 660 may use CAN and / or bus 602 (e.g., to transmit data generated by one or more RADAR sensors 660) to control and access object tracking data, and in some examples, Ethernet may be accessed to access raw data. Various RADAR sensor types may be used. For example, but not limited to, one or more RADAR sensors 660 may be suitable for front, rear, and side RADAR use. In some examples, one or more pulse Doppler RADAR sensors are used.

[0184] One or more RADAR sensors 660 may include different configurations, such as long distance with narrow field of view, short distance with wide field of view, short distance side coverage, etc. In some examples, long-range RADAR can be used for adaptive cruise control functions. The long-range RADAR system can provide a wide field of view achieved by two or more independent scans, for example within a range of 250 meters. One or more RADAR sensors 660 can help distinguish between stationary and moving objects and can be used by ADAS systems for emergency brake assistance and forward collision warnings. The long-range RADAR sensor may include a single-mode multimode radar with multiple (e.g., six or more) fixed RADAR antennas and high-speed CAN and FlexRay interfaces. In an example with six antennas, the central four antennas can create a focused beam pattern designed to record the surroundings of the vehicle 138 at higher speeds with minimal interference from traffic in adjacent lanes. The other two antennas can expand the field of view so that vehicles entering or leaving the vehicle 138 lane can be quickly detected.

[0185] As an example, a mid-range RADAR system may include a range of up to 660m (front) or 80m (rear), and a field of view of up to 42 degrees (front) or 650 degrees (rear). A short-range radar system may include, but is not limited to, a RADAR sensor designed to be mounted at both ends of the rear bumper. When mounted at both ends of the rear bumper, such a RADAR sensor system may generate two beams that continuously monitor the rear of the vehicle and the blind spots beside the vehicle.

[0186] Short-range RADAR systems can be used in ADAS systems for blind spot detection and / or lane change assistance.

[0187] The vehicle 138 may also include one or more ultrasonic sensors 662. The one or more ultrasonic sensors 662 may be placed on the front, rear, and / or sides of the vehicle 138 and may be used for parking assistance and / or creating and updating an occupancy grid. A variety of one or more ultrasonic sensors 662 may be used, and different one or more ultrasonic sensors 662 may be used for different detection ranges (e.g., 2.5m, 4m). The one or more ultrasonic sensors 662 may operate at a functional safety level of ASIL B.

[0188] The vehicle 138 may include one or more LIDAR sensors 664. The one or more LIDAR sensors 664 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The one or more LIDAR sensors 664 may be functional safety level ASIL B. In some examples, the vehicle 138 may include multiple LIDAR sensors 664 (e.g., two, four, six, etc.) that may use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0189] In some examples, one or more LIDAR sensors 664 may be able to provide a list of objects and their distances with a 360-degree field of view. One or more commercial LIDAR sensors 664 may have an advertised range of approximately 138m, an accuracy of 2cm-3cm, and support, for example, a 138Mbps Ethernet connection. In some examples, one or more non-obtrusive LIDAR sensors 664 may be used. In such examples, the LIDAR sensor 664 may be implemented as a small device that can be embedded in the front, rear, side, and / or corner of the vehicle 138. In such an example, one or more LIDAR sensors 664, may provide a horizontal field of view of up to 620 degrees and a vertical field of view of 35 degrees, providing a range of 200 meters even for low reflectivity objects. One or more front-facing LIDAR sensors 664 may be configured for a horizontal field of view between 45 degrees and 135 degrees.

[0190] In some examples, LIDAR technology, such as 3D flash LIDAR, may also be used. 3D flash LIDAR uses a laser flash as the emission source, illuminating approximately 200 meters around the vehicle. The flash LIDAR unit includes a receiver that records the laser pulse propagation time and reflected light on each pixel, which in turn corresponds to the range from the vehicle to the object. Flash LIDAR can allow each laser flash to produce a highly accurate and distortion-free image of the surrounding environment. In some examples, four flash LIDAR sensors can be deployed, one on each side of the vehicle 138. Available 3D flash LIDAR systems include solid-state 3D staring array LIDAR cameras with no moving parts other than the fan (e.g., non-scanning LIDAR devices). The flash LIDAR device can use 5 nanosecond Class I (eye-safe) laser pulses per frame, and can capture the reflected laser light in the form of a 3D range point cloud and co-registered intensity data. By using flash LIDAR, and because the flash LIDAR is a solid-state device with no moving parts, one or more LIDAR sensors 664 may be less susceptible to motion blur, vibration, and / or shock.

[0191] The vehicle may also include one or more IMU sensors 666. In some examples, the one or more IMU sensors 666 may be located at the center of the rear axle of the vehicle 138. The one or more IMU sensors 666 may include, for example but not limited to, one or more accelerometers, one or more magnetometers, one or more gyroscopes, one or more magnetic compasses, and / or other sensor types. In some examples, such as in a six-axis application, the one or more IMU sensors 666 may include an accelerometer and a gyroscope, while in a nine-axis application, the one or more IMU sensors 666 may include an accelerometer, a gyroscope, and a magnetometer.

[0192] In some embodiments, one or more IMU sensors 666 can be implemented as a miniature high-performance GPS-aided inertial navigation system (GPS / INS) that combines micro-electromechanical systems (MEMS) inertial sensors, a high-sensitivity GPS receiver, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. Thus, in some examples, one or more IMU sensors 666 can enable the vehicle 138 to estimate heading without input from a magnetic sensor by directly observing and correlating velocity changes from GPS to one or more IMU sensors 666. In some examples, one or more IMU sensors 666 and one or more GNSS sensors 658 can be combined in a single integrated unit.

[0193] The vehicle may include one or more microphones 696 placed in and / or around the vehicle 138. Among other things, the one or more microphones 696 may be used for emergency vehicle detection and identification.

[0194] The vehicle may also include any number of camera types, including one or more stereo cameras 668, one or more wide field of view cameras 670, one or more infrared cameras 672, one or more surround cameras 674, one or more long-range and / or mid-range cameras 698, and / or other camera types. The cameras may be used to capture image data around the entire perimeter of the vehicle 138. The type of camera used depends on the embodiment and requirements of the vehicle 138, and any combination of camera types may be used to provide the necessary coverage around the vehicle 138. In addition, depending on the embodiment, the number of cameras may vary. For example, the vehicle may include six cameras, seven cameras, ten cameras, twelve cameras, and / or another number of cameras. By way of example and not limitation, the cameras may support Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Fig. 6A and Figure 6B Each camera is described in more detail.

[0195] The vehicle 138 may also include one or more vibration sensors 642. The one or more vibration sensors 642 may measure vibrations of a component of the vehicle (e.g., one or more axles). For example, changes in vibration may indicate changes in the road surface. In another example, when two or more vibration sensors 642 are used, the difference between the vibrations may be used to determine friction or slip of the road surface (e.g., when there is a difference in vibration between a powered drive shaft and a free-spinning shaft).

[0196] The vehicle 138 may include an ADAS system 638. In some examples, the ADAS system 638 may include a SoC. The ADAS system 638 may include autonomous / adaptive / automatic cruise control (ACC), cooperative adaptive cruise control (CACC), forward collision warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keeping assist (LKA), blind spot warning (BSW), rear cross traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functions.

[0197] The ACC system may use one or more RADAR sensors 660, one or more LIDAR sensors 664 and / or one or more cameras. The ACC system may include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle immediately in front of the vehicle 138 and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle in front. Lateral ACC performs distance keeping and suggests that the vehicle 138 change lanes when necessary. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0198] CACC uses information from other vehicles, which can be received from other vehicles via a network interface 624 and / or one or more wireless antennas 626 via a wireless link or indirectly through a network connection (e.g., via the Internet). A direct link can be provided by a vehicle-to-vehicle (V2V) communication link, while an indirect link can be an infrastructure-to-vehicle (I2V) communication link. Typically, the V2V communication concept provides information about the vehicle immediately ahead (e.g., a vehicle immediately ahead of vehicle 138 and in the same lane as vehicle 138), while the I2V communication concept provides information about traffic farther ahead. The CACC system can include either or both of the I2V and V2V information sources. Given information about the vehicle ahead of vehicle 138, CACC may be more reliable and it has the potential to improve traffic flow and reduce road congestion.

[0199] The FCW system is designed to warn the driver of hazards so that the driver can take corrective action. The FCW system uses a front-facing camera and / or one or more RADAR sensors 660 coupled to a dedicated processor, DSP, FPGA and / or ASIC, which is electrically coupled to driver feedback, such as a display, speaker and / or vibrating component. The FCW system may provide warnings, for example, in the form of sounds, visual warnings, vibrations and / or rapid brake pulses.

[0200] The AEB system detects an impending forward collision with another vehicle or other object, and if the driver does not take corrective action within specified time or distance parameters, the AEB system may automatically apply brakes. The AEB system may use one or more forward-facing cameras and / or one or more RADAR sensors 660 coupled to a dedicated processor, DSP, FPGA and / or ASIC. When the AEB system detects a hazard, it typically first warns the driver to take corrective action to avoid the collision, and if the driver does not take corrective action, the AEB system may automatically apply brakes in an effort to prevent, or at least mitigate, the effects of the predicted collision. The AEB system may include technologies such as dynamic brake support and / or collision approach braking.

[0201] The LDW system provides visual, audible, and / or tactile warnings, such as steering wheel or seat vibrations, to warn the driver when the vehicle 138 crosses a lane marking. The LDW system does not activate when the driver indicates intentional lane departure by activating a turn signal. The LDW system may use a front-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback, such as a display, speaker, and / or vibration component.

[0202] The LKA system is a variation of the LDW system. If the vehicle 138 begins to leave the lane, the LKA system provides steering input or braking to correct the vehicle 138.

[0203] The BSW system detects and warns the driver of vehicles in the car's blind spot. The BSW system may provide visual, audible, and / or tactile alerts to indicate that it is unsafe to merge or change lanes. The system may provide additional warnings when the driver uses a turn signal. The BSW system may use one or more rear side cameras and / or one or more RADAR sensors 660 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, which is electrically coupled to driver feedback, such as a display, speaker, and / or vibration component.

[0204] When the vehicle 138 is backing up, the RCTW system may provide visual, audible, and / or tactile notifications when an object is detected outside of the rear camera range. Some RCTW systems include AEB to ensure that the vehicle brakes are applied to avoid a collision. The RCTW system may use one or more rear RADAR sensors 660 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, which is electrically coupled to driver feedback, such as a display, speaker, and / or vibration element.

[0205] Conventional ADAS systems may be prone to false positive results, which may be annoying and distracting to the driver, but are generally not catastrophic because the ADAS system warns the driver and allows the driver to determine whether the safety condition actually exists and act accordingly. However, in the autonomous vehicle 138, in the event of conflicting results, the vehicle 138 itself must decide whether to pay attention to the results from the main computer or the auxiliary computer (e.g., the first controller 636 or the second controller 636). For example, in some embodiments, the ADAS system 638 can be a backup and / or auxiliary computer for providing perception information to the backup computer rationality module. The backup computer rationality monitor can run redundant various software on hardware components to detect failures in perception and dynamic driving tasks. The output from the ADAS system 638 can be provided to the supervisory MCU. If the outputs of the main computer and the auxiliary computer conflict, the monitoring MCU must determine how to reconcile the conflict to ensure safe operation.

[0206] In some examples, the master computer can be configured to provide a confidence score to the supervisory MCU, indicating the master computer's confidence in the selected result. If the confidence score exceeds a threshold, the supervisory MCU can follow the master computer's instructions regardless of whether the auxiliary computer provides conflicting or inconsistent results. In the event that the confidence score does not meet the threshold and the master computer and the auxiliary computer indicate different results (e.g., conflict), the supervisory MCU can arbitrate between the computers to determine the appropriate result.

[0207] The supervisory MCU can be configured to run one or more neural networks that are trained and configured to determine conditions under which the auxiliary computer provides a false alarm based on outputs from the main computer and the auxiliary computer. Therefore, one or more neural networks in the monitoring MCU can learn when the output of the auxiliary computer can be trusted and when it cannot be trusted. For example, when the auxiliary computer is a RADAR-based FCW system, one or more neural networks in the monitoring MCU can learn when the FCW system identifies a metal object that triggers an alarm that is not actually dangerous, such as a drain grate or manhole cover. Similarly, when the auxiliary computer is a camera-based LDW system, the neural network in the monitoring MCU can learn that overtaking the LDW and lane departure when there are cyclists or pedestrians is actually the safest maneuver. In an embodiment including one or more neural networks running on the monitoring MCU, the monitoring MCU may include at least one of a DLA or GPU suitable for running one or more neural networks with associated memory. In a preferred embodiment, the monitoring MCU may include and / or be included as a component of one or more SoC 604.

[0208] In other examples, the ADAS system 638 may include an auxiliary computer that uses traditional computer vision rules to perform ADAS functions. In this way, the auxiliary computer may use classical computer vision rules (hypothetical), and the presence of one or more neural networks in the supervisory MCU may improve reliability, safety, and performance. For example, diverse implementations and intentional non-identity make the entire system more fault-tolerant, especially failures caused by software (or software-hardware interface) functions. For example, if there is a software error or fault in the software running on the main computer, and a different software code running on the auxiliary computer provides the same overall result, the monitoring MCU may have more confidence that the overall result is correct and that the error in the software or hardware on the main computer did not cause a gross fault.

[0209] In some examples, the output of the ADAS system 638 can be fed to the perception block of the main computer and / or the dynamic driving task block of the main computer. For example, if the ADAS system 638 indicates a forward collision warning due to an object immediately ahead, the perception block can use this information in identifying the object. In other examples, as described herein, the auxiliary computer can have its own trained neural network, thereby reducing the risk of false positives.

[0210] The vehicle 138 may also include an infotainment system SoC 630 (e.g., an in-vehicle infotainment system (IVI)). Although shown and described as an SoC, the infotainment system may not be a SoC and may include two or more discrete components. The infotainment system SoC 630 may include a combination of hardware and software that may be used to provide audio (e.g., music, personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., television, movies, streaming, etc.), phone (e.g., hands-free calling), network connectivity (e.g., LTE, WiFi, etc.), and / or information services (e.g., navigation system, rear parking assist, radio data system, vehicle-related information such as fuel level, total distance covered, brake fuel level, oil level, door open / closed, air filter information, etc.) to the vehicle 138. For example, the infotainment system SoC 630 can be a radio, a disk player, a navigation system, a video player, USB and Bluetooth connectivity, onboard, in-car entertainment, WiFi, steering wheel audio controls, hands-free voice control, a head-up display (HUD), an HMI display 634, a telematics device, a control panel (e.g., for controlling and / or interacting with various components, features, and / or systems), and / or other components. The infotainment system SoC 630 can also be used to provide information (e.g., visual and / or auditory) to one or more users of the vehicle, such as information from the ADAS system 638, such as autonomous driving information of planned vehicle maneuvers, trajectories, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0211] The infotainment system SoC 630 may include GPU functionality. The infotainment system SoC 630 may communicate with other devices, systems, and / or components of the vehicle 138 via a bus 602 (e.g., CAN bus, Ethernet, etc.). In some examples, the infotainment system SoC 630 may be coupled to a supervisory system MCU so that the GPU of the infotainment system may perform some self-driving functions in the event of a failure of one or more host controllers 636 (e.g., the primary and / or backup computers of the vehicle 138). In such an example, the infotainment system SoC 630 may place the vehicle 138 in a driver safety stop mode, as described herein.

[0212] The vehicle 138 may also include an instrument cluster 632 (e.g., a digital instrument panel, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 632 may include a controller and / or a supercomputer (e.g., a discrete controller or a supercomputer). The instrument cluster 632 may include a set of instruments, such as a speedometer, a fuel level, an oil pressure, a tachometer, an odometer, a turn indicator, a shift position indicator, one or more seat belt warning lights, one or more parking brake warning lights, one or more parking brake function lights, one or more engine fault lights, airbag (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some examples, information may be displayed and / or shared between the infotainment system SoC 630 and the instrument cluster 632. In other words, the instrument cluster 632 may be included as part of the infotainment system SoC 630, and vice versa.

[0213] Fig.6D According to some embodiments of the present disclosure, Fig. 6A 6. System diagram of exemplary autonomous vehicle 138 and one or more cloud-based servers. System 676 may include one or more servers 678, one or more networks 690, and multiple vehicles including vehicle 138. One or more servers 678 may include multiple GPUs 684(A)-684(H) (collectively referred to herein as GPUs 684), PCIe switches 682(A)-682(H) (collectively referred to herein as PCIe switches 682), and / or CPUs 680(A)-680(B) (collectively referred to herein as CPUs 680). GPUs 684, CPUs 680, and PCIe switches may be interconnected with a high-speed interconnect such as, but not limited to, NVLink interface 688 developed by NVIDIA and / or PCIe connection 686. In some examples, GPU 684 is connected via NVLink and / or NVSwitch SoC, and GPU 684 and PCIe switch 682 are connected via a PCIe interconnect. Although eight GPUs 684, two CPUs 680, and two PCIe switches are shown, this is not limiting. Depending on the embodiment, each server 678 may include any number of GPUs 684, CPUs 680, and / or PCIe switches. For example, one or more servers 678 may each include eight, sixteen, thirty-two, and / or more GPUs 684.

[0214] One or more servers 678 may receive image data representing an image of an unexpected or changed road condition (e.g., a recently begun road project) from a vehicle via one or more networks 690. One or more servers 678 may send information including information about traffic and road conditions to the vehicle, neural network 692, updated neural network 692, and / or map information 694 via one or more networks 690. Updates to map information 694 may include updates to HD map 622, such as information about construction sites, potholes, detours, flooding, and / or other obstacles. In some examples, neural network 692, updated neural network 692, and / or map information 694 may be generated by new training and / or experience represented in data received from any number of vehicles in the environment, and / or based on training performed at a data center (e.g., using one or more servers 678 and / or other servers).

[0215] One or more servers 678 can be used to train a machine learning model (e.g., a neural network) based on training data. The training data can be generated by the vehicle, and / or can be generated in simulation (e.g., using a game engine). In some examples, the training data is labeled (e.g., neural networks benefit from supervised learning) and / or subjected to other preprocessing, while in other examples, the training data is not labeled and / or preprocessed (e.g., neural networks do not require supervised learning). Once the machine learning model is trained, the machine learning model can be used by the vehicle (e.g., transmitted to the vehicle via one or more networks 690, and / or the machine learning model can be used by one or more servers 678) to remotely monitor the vehicle.

[0216] In some examples, one or more servers 678 can receive data from the vehicle and apply the data to the latest real-time neural network for real-time intelligent reasoning. One or more servers 678 may include a deep learning supercomputer and / or a dedicated AI computer driven by one or more GPUs 684 (e.g., DGX and DGX Station machines developed by NVIDIA). However, in some examples, one or more servers 678 may include a deep learning infrastructure of a data center powered only by CPUs.

[0217] The deep learning infrastructure of one or more servers 678 may be capable of fast, real-time inference, and may use this capability to assess and validate the health of the processor, software, and / or associated hardware. For example, the deep learning infrastructure may receive periodic updates from the vehicle 138, such as a series of images and / or objects in which the vehicle 138 has been located (e.g., via computer vision and / or other machine learning object classification techniques). The deep learning infrastructure may run its own neural network to identify objects and compare them to the objects identified by the vehicle 138, and if the results do not match and the infrastructure concludes that the AI ​​in the vehicle 138 has failed, the one or more servers 678 may send a signal to the vehicle 138 instructing the fail-safe computer of the vehicle 138 to take control, notify the passengers, and complete a safe parking maneuver.

[0218] For inference, one or more servers 678 may include a GPU 684 and one or more programmable inference accelerators (e.g., NVIDIA's TensorRT). The combination of GPU-powered servers and inference acceleration can make real-time responses possible. In other examples, such as when performance is less important, servers powered by CPUs, FPGAs, and other processors can be used for inference.

[0219] Example computing device

[0220] Figure 7 is a block diagram of an example computing device 700 suitable for implementing some embodiments of the present disclosure. The computing device 700 may include a bus 702 that directly or indirectly couples the following devices: a memory 704, one or more central processing units (CPUs) 706, one or more graphics processing units (GPUs) 708, a communication interface 710, an input / output (I / O) port 712, a plurality of input / output components 714, a power supply 716, and one or more presentation components 718 (e.g., one or more displays).

[0221] although Figure 7 The various blocks are shown connected to the bus 702 via lines, which is not limiting and is for clarity only. For example, in some embodiments, a presentation component 718 such as a display device may be considered an I / O component 714 (e.g., if the display is a touch screen). As another example, the CPU 706 and / or the GPU 708 may include a memory (e.g., the memory 704 may represent a storage device in addition to the memory of the GPU 708, the CPU 706, and / or other components). In other words, Figure 7The computing devices in the description are illustrative only. No distinction is made between categories such as "workstation," "server," "laptop," "desktop," "tablet," "client device," "mobile device," "handheld device," "gaming console," "electronic control unit (ECU)," "virtual reality system," and / or other device or system types, as in Figure 7 as expected within the range of computing devices.

[0222] The bus 702 may represent one or more buses, such as an address bus, a data bus, a control bus, or a combination thereof. The bus 702 may include one or more bus types, such as an Industry Standard Architecture (ISA) bus, an Extended Industry Standard Architecture (EISA) bus, a Video Electronics Standards Association (VESA) bus, a Peripheral Component Interconnect (PCI) bus, a Peripheral Component Interconnect Express (PCIe) bus, and / or other types of buses.

[0223] Memory 704 may include any of a variety of computer-readable media. Computer-readable media may be any available media that can be accessed by computing device 700. Computer-readable media may include volatile and non-volatile media, as well as removable and non-removable media. By way of example and not limitation, computer-readable media may include computer storage media and communication media.

[0224] Computer storage media may include volatile and nonvolatile media and / or removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, and / or other data types. For example, memory 704 may store computer readable instructions (e.g., representing one or more programs and / or one or more program elements, such as an operating system). Computer storage media may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other storage technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by the computing device 700. As used herein, computer storage media itself does not include signals.

[0225] Communication media may embody computer readable instructions, data structures, program modules and / or other data types in a modulated data signal such as a carrier wave or other transport mechanism, and includes any information delivery media. The term "modulated data signal" may refer to a signal that has one or more of its characteristics set or changed in a manner that encodes the information in the signal. By way of example and not limitation, communication media may include wired media, such as a wired network or direct wired connection, and wireless media, such as acoustic, RF, infrared and other wireless media. Any combination of the above should also be included within the scope of computer readable media.

[0226] One or more CPUs 706 may be configured to execute computer-readable instructions to control one or more components of the computing device 700 to perform one or more of the methods and / or processes described herein. Each CPU 706 may include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of processing multiple software threads simultaneously. One or more CPUs 706 may include any type of processor, and may include different types of processors, depending on the type of computing device 700 implemented (e.g., a processor with fewer cores for mobile devices and a processor with more cores for servers). For example, depending on the type of computing device 700, the processor may be an ARM processor implemented using reduced instruction set computing (RISC) or an x86 processor implemented using complex instruction set computing (CISC). In addition to one or more microprocessors or auxiliary coprocessors (e.g., math coprocessors), the computing device 700 may also include one or more CPUs 706.

[0227] The computing device 700 may use one or more GPUs 708 to render graphics (e.g., 3D graphics). One or more GPUs 708 may include hundreds or thousands of cores that can process hundreds or thousands of software threads simultaneously. One or more GPUs 708 may generate pixel data of an output image in response to a rendering command (e.g., a rendering command rendered from one or more CPUs 706 via a host interface). One or more GPUs 708 may include a graphics memory for storing pixel data, such as a display memory. Display memory may be included as part of the memory 704. One or more GPUs 708 may include two or more GPUs operating in parallel (e.g., via a link). When combined together, each GPU 708 may generate pixel data for different parts of an output image or for different output images (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU may include its own memory, or may share memory with other GPUs.

[0228] In examples where computing device 700 does not include one or more GPUs 708, one or more CPUs 706 may be used to render graphics.

[0229] The communication interface 710 may include one or more receivers, transmitters, and / or transceivers that enable the computing device 700 to communicate with other computing devices via an electronic communication network (including wired and / or wireless communications). The communication interface 710 may include components and functionality for enabling communication over any of a number of different networks, such as a wireless network (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, etc.), a wired network (e.g., communicating over Ethernet), a low power wide area network (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0230] I / O ports 712 may enable computing device 700 to be logically coupled to other devices including I / O components 714, one or more presentation components 718, and / or other components, some of which may be built into (e.g., integrated into) computing device 700. Exemplary I / O components 714 include microphones, mice, keyboards, joysticks, gamepads, game controllers, satellite dishes, scanners, printers, wireless devices, and the like. I / O components 714 may provide a natural user interface (NUI) that processes air gestures, speech, or other physiological input generated by a user. In some cases, the input may be sent to an appropriate network element for further processing. The NUI may implement any combination of speech recognition, stylus recognition, facial recognition, biometrics, gesture recognition on and near the screen, air gestures, head and eye tracking, and touch recognition associated with the display of computing device 700 (as described in more detail below). Computing device 700 may include a depth camera, such as a stereo camera system, an infrared camera system, an RGB camera system, touch screen technology, and combinations thereof, for gesture detection and recognition. Additionally, computing device 700 may include an accelerometer or gyroscope capable of detecting motion (e.g., as part of an inertial measurement unit (IMU)). In some examples, the output of the accelerometer or gyroscope may be used by computing device 700 to present an immersive augmented reality or virtual reality.

[0231] The power supply 716 may include a hardwired power supply, a battery power supply, or a combination thereof. The power supply 716 may provide power to the computing device 700 to enable the components of the computing device 700 to operate.

[0232] The one or more presentation components 718 may include a display (e.g., a monitor, a touch screen, a television screen, a heads-up display (HUD), other display types, or combinations thereof), speakers, and / or other presentation components. The one or more presentation components 718 may receive data from other components (e.g., one or more GPUs 708, one or more CPUs 706, etc.) and output data (e.g., as images, video, sound, etc.).

[0233] The present disclosure may be described in the general context of computer code or machine usable instructions, including computer executable instructions, such as program modules, executed by a computer or other machine (such as a personal data assistant or other handheld device). In general, program modules, including routines, programs, objects, components, data structures, etc., refer to code that performs specific tasks or implements specific abstract data types. The present disclosure may be practiced in a variety of system configurations, including handheld devices, consumer electronics, general-purpose computers, more specialized computing devices, etc. The present disclosure may also be implemented in a distributed computing environment, where tasks are performed by remote execution-processing devices linked through a communications network.

[0234] As used herein, the statement of "and / or" about two or more elements should be interpreted as indicating only one element or a combination of elements. For example, "element A, element B and / or element C" may include only element A, only element B, only element C, element A and element B, element A and element C, element B and element C, or elements A, B and C. In addition, "at least one of element A or element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. In addition, "at least one of element A and element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B.

[0235] The subject matter of the present disclosure is described in detail herein to meet statutory requirements. However, the description itself is not intended to limit the scope of the present disclosure. Instead, the inventors have conceived that the claimed subject matter may also be embodied in other ways, to include different steps or combinations of steps similar to the steps described in this document, in combination with other existing or future technologies. In addition, although the terms "step" and / or "block" may be used herein to refer to different elements of the method employed, these terms should not be interpreted as implying any particular order among or between the various steps disclosed herein, unless the order of the various steps is explicitly described.

Claims

1. A computer-implemented method comprising: receiving trajectory training data for at least one vehicle and sensor data representing a sensing field of a sensor of the at least one vehicle; Trajectories are generated by performing the following operations for each representation of the sensor data: generating a trajectory for the vehicle using a subset of the trajectory training data corresponding to a path of the vehicle within at least a portion of the sensing field of the sensor in the representation; as well as associating the trajectory with the representation of the sensor data by using the trajectory as a label or annotation corresponding to the representation of the sensor data; applying the sensor data as input to a machine learning model; as well as The machine learning model is trained based at least in part on comparing outputs computed by the machine learning model to the trajectory as ground truth data. 2 . The method of claim 1 , wherein the trajectory training data comprises, at least in part, trajectory data automatically labeled based on associating perception sensor data with speed and orientation data generated during a human-driven portion of vehicle operation. 3 . The method of claim 1 , wherein the trajectory training data at least partially comprises training data annotated in 2D image space and converted to trajectory labels in 3D world coordinates. The method of claim 1 , wherein comparing the output to the trajectory is performed using a loss function. The method of claim 1 , wherein the trajectory is represented by one or more trajectory points, each of the one or more trajectory points representing a three-dimensional coordinate in a world space. 6 . The method of claim 5 , wherein a first trajectory of the trajectory is represented by a plurality of first trajectory points, and a distance between each of the plurality of first trajectory points comprises a static distance in world space.

7. The method of claim 5 , wherein a first trajectory in the trajectories is represented by a plurality of first trajectory points, and a distance between a particular first trajectory point in the plurality of first trajectory points and at least one of a previous first trajectory point or a next first trajectory point in the plurality of first trajectory points is based on a distance between the particular first trajectory point and the vehicle.

8. The method of claim 5 , wherein a first trajectory of the trajectory is represented by a plurality of first trajectory points within a threshold distance from the vehicle and a plurality of second trajectory points beyond the threshold distance from the vehicle, a distance between each of the plurality of first trajectory points is a first static distance, and a distance between each of the plurality of second trajectory points is a second static distance, the second static distance being farther than the first static distance.

9. The method according to claim 1, further comprising: Generate transformed sensor data for each representation of the sensor data: determining a first perspective of the sensor generating the sensor data; and converting the sensor data into a second perspective of a second sensor; Wherein applying the sensor data as the input to the machine learning model comprises applying the transformed sensor data.

10. The method of claim 9, wherein the converting to the second perspective comprises at least one of shifting the sensor data or rotating the sensor data.

11. The method of claim 1 , wherein the sensor comprises one or more of the following sensors: One or more sensor devices, including at least one of the following: Image sensor; LIDAR sensor; Radar sensor; Ultrasonic sensor; or Sonar sensors; Global Positioning System (GPS) devices; Inertial Measurement Unit (IMU); or Speedometer.

12. A method according to claim 1, wherein the machine learning model is further trained using additional ground truth data, the additional ground truth data comprising annotated tracks associated with training images represented by the sensor data, the annotated tracks automatically generated at least in part based on speed data from one or more speed sensors of the vehicle and orientation data from one or more inertial measurement unit sensors of the vehicle.

13. A method according to claim 12, wherein a training image of a training image is generated at a first time, the speed data and the orientation data associated with an annotation track of the training image are generated one or more times after the first time, and the annotation track represents the track of the data acquisition vehicle within the field of view of the data acquisition camera that captured the training image.

14. A computer-implemented method comprising: receiving sensor data representing a sensing field of a sensor disposed or positioned on a vehicle moving through an environment; Computing trajectory data representing a world-space trajectory of the vehicle using one or more machine learning models and based at least on the sensor data, the one or more machine learning models being trained based on: ground truth data representing one or more trajectories for navigating a data collection vehicle within a real-world environment; image data generated along the one or more tracks within the real-world environment during navigation of the data collection vehicle using one or more image sensors of the data collection vehicle, the image data representing one or more images depicting at least a portion of the real-world environment; one or more annotations associating the ground truth data with the image data, a single annotation of the one or more annotations associating a position along one of the one or more tracks with one of the one or more images; determining a predicted trajectory using the one or more machine learning models and based on at least a portion of the image data representing the image; as well as comparing the predicted trajectory to at least a portion of the one or more trajectories used to navigate a data collection vehicle within the real-world environment based at least on the one or more annotations; as well as The vehicle is caused to navigate based at least on the world-space trajectory.

15. A method according to claim 14, wherein the one or more machine learning models are further trained using training data of trajectory labels annotated in two-dimensional 2D image space and converted to three-dimensional 3D world space coordinates, and the 3D world space coordinates are used as ground truth outputs of the one or more machine learning models.

16. The method of claim 14, wherein the world-space trajectory comprises one or more trajectory points, a control component uses at least at least one trajectory point of the one or more trajectory points to generate a control profile for the vehicle, and the control profile comprises: At least one of accelerating, decelerating or turning.

17. A computer-implemented method comprising: receiving image data generated by a first camera having first camera parameters, the first camera being associated with a vehicle in an environment; applying a transformation to the image data to generate transformed image data, the transformation converting the image data to a perspective of a second camera having second camera parameters; applying the transformed image data to a machine learning model trained using training image data representing training images generated from the view of the second camera; computing, by the machine learning model, output data representing one or more trajectory points in world space and representing a recommended trajectory for the vehicle through the environment; and The output data is sent to a control component of the vehicle to cause the vehicle to perform one or more operations based at least in part on the output data.

18. The method of claim 17, wherein the output data further represents the orientation of the vehicle relative to one or more features of the environment.

19. The method of claim 18, wherein the one or more features include at least one of a lane marking, a driving surface boundary, or another vehicle.

20. The method of claim 17, wherein the output data further represents a state of the vehicle relative to one or more vehicle maneuvers.

21. The method of claim 20, wherein the one or more vehicle maneuvers include at least one of a turn, a lane change, a lane merge, a lane split, or a road exit.

Images