Deep Neural Networks for Obstacle Instance Detection Using RADAR Sensors in Autonomous Machine Applications

Abstract

To provide object detection for autonomous machines using deep neural networks (DNNs).SOLUTION: A DNN is trained to detect moving and stationary obstacles from RADAR data of a 3D space. Ground truth training data for the DNN is generated from LIDAR data. More specifically, a scene is observed with RADAR and LIDAR sensors to collect RADAR and LIDAR data for a particular time slice. The RADAR data is used for input training data, and the LIDAR data associated with the same time slice as or the closest time slice to the RADAR data is annotated with ground truth labels identifying objects to be detected. The LIDAR labels are propagated to the RADAR data, and LIDAR labels containing less than a threshold number of RADAR detection are omitted. The remaining LIDAR labels are used to generate ground truth data.SELECTED DRAWING: Figure 1

Description

[Technical field]

[0001] This application claims the benefit of U.S. Provisional Patent Application No. 62 / 938,852, filed Nov. 21, 2019, which is incorporated by reference in its entirety. [Background technology]

[0002] Designing a system that safely drives a vehicle autonomously without supervision is extremely challenging. An autonomous vehicle should at least be capable of performing as the functional equivalent of a careful driver, utilizing a perception and action system with an extraordinary ability to identify and react to moving and stationary obstacles in a complex environment to avoid collisions with other objects or structures along the vehicle's path. Thus, the ability to detect instances of moving or stationary actors (e.g., passenger vehicles, pedestrians, etc.) is a critical component of an autonomous driving perception system. This capability becomes increasingly important as the operating environment of autonomous vehicles begins to expand from highway environments to semi-urban and urban settings characterized by complex scenes with numerous occlusions and complex geometries.

[0003] Traditional perception methods rely heavily on the use of cameras or LIDAR sensors to detect obstacles in a scene. However, these traditional approaches have several drawbacks. For example, traditional detection techniques are unreliable in scenes with heavy occlusions. In addition, traditional sensing techniques are generally unreliable in adverse weather conditions, and basic sensors are often prohibitively expensive. Furthermore, the output signals from these traditional systems require extensive post-processing to extract accurate three-dimensional (3D) information, so the runtime of these traditional systems is generally more advanced, requiring additional computation and processing charges, thereby reducing the efficiency of these traditional systems.

[0004] Some conventional techniques use RADAR sensors to detect moving reflective objects. However, many conventional RADAR detection techniques struggle or fail completely to separate obstacles from background noise in cluttered environments. Furthermore, while some conventional RADAR detection techniques perform well when detecting moving RADAR reflective objects, these conventional techniques often struggle or fail completely to distinguish stationary objects from background noise. Similarly, conventional RADAR detection techniques have limited accuracy in predicting object classification, dimensions, and orientation. [Prior art documents] [Patent documents]

[0005] [Patent Document 1] U.S. Patent Application Serial No. 16 / 101,232 Summary of the Invention [Means for solving the problem]

[0006] FIELD OF THE DISCLOSURE Embodiments of the present disclosure relate to object detection for autonomous machines using deep neural networks (DNNs). Systems and methods are disclosed that use object detection techniques to identify or detect instances of moving or stationary obstacles (e.g., cars, trucks, pedestrians, cyclists, etc.) and other objects in an environment for use by autonomous vehicles, semi-autonomous vehicles, robots, and / or other object types.

[0007] In contrast to conventional systems such as those described above, the disclosed system can implement a deep learning solution (e.g., using a deep neural network (DNN) such as a convolutional neural network (CNN)) in an autonomous vehicle to detect moving and stationary obstacles and other objects from RADAR data. More specifically, the neural network can be trained to detect moving and stationary obstacles from RADAR data in three-dimensional (3D) space. The RADAR detections can be accumulated, ego-motion corrected, orthogonally projected, and fed into the neural network. The neural network can include a common trunk with a feature extractor and several heads that predict different outputs, such as a class confidence head that predicts a confidence map of "present" objects and an instance regression head that predicts object instance data (e.g., location, dimensions, pose, orientation, etc.) of the detected objects. The outputs can be decoded, filtered, and / or clustered to form a bounding shape that identifies the location, size, and / or orientation of the detected object instances. The detected object instances can be provided to the autonomous machine control stack to enable safe planning and control of the autonomous machine.

[0008] In some embodiments, ground truth training data for a neural network may be generated from LIDAR data. More specifically, a scene may be observed with RADAR and LIDAR sensors to collect RADAR and LIDAR data for a particular time slice. The RADAR data may be used for input training data, and LIDAR data associated with the same or closest time slice as the RADAR data may be annotated with ground truth labels that identify objects to be detected. The LIDAR labels may be propagated to the RADAR data, and LIDAR labels with fewer than some threshold number of RADAR detections may be omitted. The (remaining) LIDAR labels may be used to generate the ground truth data. As such, the training data may be used to train a DNN to detect moving and stationary obstacles and other objects from the RADAR data.

[0009] Unlike conventional approaches, the present technique can be used to distinguish between stationary obstacles such as vehicles and stationary background noise, which is especially important when navigating cluttered urban environments. Furthermore, because embodiments of the present disclosure can rely on RADAR data during operation, object detection is feasible in adverse weather and at night, situations where camera-based and LIDAR-based detection techniques degrade and fail.

[0010] The present system and method for detecting obstacle instances using a RADAR sensor in autonomous machine applications is described in detail below with reference to the accompanying drawings. [Brief description of the drawings]

[0011] [Figure 1] 1 is a data flow diagram illustrating an example process of an object detection system according to some embodiments of the present disclosure. [Diagram 2]1 is a data flow diagram illustrating an example process for preprocessing sensor data for a machine learning model in an object detection system, in accordance with some embodiments of the present disclosure. [Diagram 3] FIG. 2 illustrates an example machine learning model, according to some embodiments of the present disclosure. [Figure 4] 1 is a data flow diagram illustrating an example post-processing process for generating object detections in an object detection system, according to some embodiments of the present disclosure. [Figure 5A] FIG. 2 illustrates an example orthogonal projection of accumulated RADAR detections and corresponding object detections, in accordance with some embodiments of the present disclosure. [Figure 5B] FIG. 1 is a diagram of object detection projected onto corresponding images from three cameras, according to some embodiments of the present disclosure. [Figure 6] 1 is a flow diagram illustrating a method for object detection using RADAR data, according to some embodiments of the present disclosure. [Figure 7] 1 is a flow diagram illustrating another method for object detection using RADAR data, according to some embodiments of the present disclosure. [Figure 8] 1 is a flow diagram illustrating another method for object detection using RADAR data, according to some embodiments of the present disclosure. [Figure 9A] FIG. 13 illustrates example orthogonal projections of accumulated RADAR detections and corresponding ground truth bounding boxes propagated from corresponding LIDAR frames and filtered based on a minimum number of RADAR detections per bounding box, in accordance with some embodiments of the present disclosure. [Figure 9B] FIG. 1 is a diagram of a corresponding LIDAR frame with ground truth bounding boxes, in accordance with some embodiments of the present disclosure. [Figure 10] 1 is a flow diagram illustrating a method for training a machine learning model for object detection according to some embodiments of the present disclosure. [Figure 11] 1 is a flow diagram illustrating another method for training a machine learning model for object detection according to some embodiments of the present disclosure. [Figure 12] 1 is a flow diagram illustrating a method for conveying LIDAR labels to RADAR data, according to some embodiments of the present disclosure. [Figure 13A] 1 is an illustration of an example autonomous vehicle, according to some embodiments of the present disclosure. [Figure 13B] 13B is an illustration of camera positions and fields of view for the example autonomous vehicle of FIG. 13A, according to some embodiments of the present disclosure. [Figure 13C] FIG. 13B is a block diagram of an example system architecture of the example autonomous vehicle of FIG. 13A in accordance with some embodiments of the present disclosure. [Figure 13D] FIG. 13B is a system diagram of communication between a cloud-based server and the example autonomous vehicle of FIG. 13A, according to some embodiments of the present disclosure. [Figure 14] FIG. 1 is a block diagram of an example computing device suitable for use in implementing some embodiments of the present disclosure. DETAILED DESCRIPTION OF THE PREFERRED EMBODIMENTS

[0012] Systems and methods related to object detection for autonomous machines using deep neural networks (DNNs) are disclosed. Systems and methods are disclosed that use object detection techniques to identify or detect instances of moving or stationary obstacles (e.g., cars, trucks, pedestrians, cyclists, etc.) and other objects in an environment for use by autonomous vehicles, semi-autonomous vehicles, robots, and / or other object types. The present disclosure is described with respect to an exemplary autonomous vehicle 1300 (an example of which is described herein with respect to FIGS. 13A-13D and alternatively referred to herein as “vehicle 1300” or “ego vehicle 1300”), but this is not intended to be limiting. For example, the systems and methods described herein may 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 ships, boats, and / or other vehicle types. Additionally, although the present disclosure may be described with respect to autonomous driving, this is not intended to be limiting. For example, the systems and methods described herein may be used in other technical fields, such as robotics (e.g., path planning for robots), aviation systems (e.g., path planning for drones or other aircraft), marine systems (e.g., path planning for boats or other water vessels), and / or localization, path planning, and / or other processes.

[0013] At a high level, a DNN (e.g., a Convolutional Neural Network (CNN)) can be trained to detect moving and stationary obstacles using RADAR data in three-dimensional (3D) space in both highway and urban scenarios. To form inputs to the DNN, raw RADAR detections of the environment around an ego object or ego actor (e.g., a moving vehicle) can be preprocessed into a format that the DNN understands. Specifically, the RADAR detections can be accumulated, transformed to a single coordinate system (e.g., centered on the ego actor), ego-motion corrected (e.g., to the last known location of the ego actor), and / or orthogonally projected to form a projected image (e.g., an overhead image) of a desired size (e.g., spatial dimensions) and with a desired ground sampling distance. For each pixel of the projection image that obtains one or more detections, a set of features may be calculated or determined from the reflectance characteristics of the RADAR detections (e.g., bearing, azimuth, altitude, range, brightness, Doppler velocity, RADAR cross section (RCS), reflectivity, signal-to-noise ratio (SNR), etc.). When there are multiple detections obtained on (e.g., across) the pixel, a particular feature may be calculated by aggregating (e.g., using standard deviation, average, etc.) the corresponding reflectance characteristics of the multiple overlapping detections. Thus, any given pixel may have multiple associated feature values ​​that may be stored in corresponding channels of the tensor. As such, the RADAR detections may be preprocessed into a multi-channel RADAR data tensor of a desired size, where each pixel of the projection image contained therein may include an associated set of feature values ​​generated from the accumulated and / or ego-motion corrected RADAR detections. This RADAR data tensor may serve as an input to the DNN.

[0014] The architecture of the DNN can allow features to be extracted from the RADAR data tensors and class classification and / or instance regression can be performed on the extracted features. For example, the DNN can include a common trunk (or stream of layers) with several heads (or at least partially discrete streams of layers) for predicting different outputs based on the input data. In an embodiment (e.g., similar to a feature pyramid network, U-Net, etc.), the common trunk can be implemented using encoder and decoder components with skip connections. The output of the common trunk can be connected to a class confidence head and / or an instance regression head. To help predict classification data (e.g., confidence map) in the form of a multi-channel tensor, the class confidence head can include a channel (e.g., classifier) ​​for each class of object to be detected (e.g., vehicle, car, truck, vulnerable road user, pedestrian, cyclist, motorcycle, etc.). Each channel can be considered as a heat map with a confidence / likelihood value that each pixel belongs to the class corresponding to the channel. The instance regression head may include N channels (e.g., classifiers), where each channel regresses a particular type of information about a detected object, such as where the object is located (e.g., a dx / dy vector pointing to the center of the object), the object height, the object width, the object orientation (e.g., rotation angles such as sine and / or cosine), and / or the like. Thus, the instance regression head can help predict a multi-channel instance regression tensor that stores N types of object information. Each channel of the instance regression tensor may include a floating point number that regresses a particular type of object information, such as a particular object dimension. As a non-limiting example, each pixel of the instance regression tensor may be:<dx、dy、w、h、sinO、cosOなど> As such, the DNN can predict a multi-channel class confidence tensor and / or a multi-channel instance regression tensor from a given RADAR data tensor.

[0015] The predicted class confidence tensor and the instance regression tensor may be used to generate a bounding box, closed polyline, or other bounding shape that identifies the position, size, and / or orientation of the detected object instance within the scene shown in the projection image. Because the object instance data may be noisy, the bounding shape may be generated using non-maximum suppression, density-based spatial clustering of application with noise (DBSCAN), and / or another function. As a non-limiting example, a candidate bounding box (or other bounding shape) may be formed for a given object class based on the object instance data (e.g., dimensions such as position, size, pose, and / or orientation data, etc.) from the corresponding channel of the instance regression tensor and / or from the confidence map from the corresponding channel of the class confidence tensor for that class. The result may be a set of candidate bounding boxes (or other bounding shapes) for each object class.

[0016] Various types of filtering may be performed to remove certain candidates. For example, each candidate may be associated with a corresponding confidence / likelihood value associated with one or more corresponding pixels from a corresponding channel of the class confidence tensor of the class being evaluated (e.g., using the confidence / likelihood value of a representative pixel such as the central pixel, using an average or some other composite value calculated over the candidate region, etc.). Thus, candidate boundary shapes with confidence / likelihood of being a member of the object class below some threshold (e.g., 50%) may be filtered out. The candidate with the highest confidence / likelihood score for the class may be assigned an instance ID, a metric such as intersection over union (IoU) may be calculated for each of the other candidates in the class, and candidates with IoU above some threshold may be filtered out to remove duplicates. The process may then be repeated, assigning an instance ID to the candidate with the highest confidence / likelihood score, removing duplicates, and so on until no candidates remain. This process may be repeated for each of the other classes. Additionally and / or alternatively, clustering may be performed on the candidate bounding shapes, for example, by clustering around the centers of the candidate bounding shapes and removing overlapping candidates from each cluster.

[0017] As such, post-processing may be applied to the predicted class confidence tensors and instance regression tensors to generate bounding boxes, closed polylines, or other bounding shapes that identify the position, size, and / or orientation of the detected object instance within the scene shown in the corresponding projection image. Once the object instance is determined, the 2D pixel coordinates that define the object instance may be transformed into 3D world coordinates (e.g., by projecting the detected object instance from a 2D orthogonal projection back to the 3D world coordinates) for use by the autonomous vehicle in performing one or more operations (e.g., lane keeping, lane changing, route planning, mapping, etc.).

[0018] To train the DNN, training data may be generated using the pre-processing techniques described above. However, given how sparse RADAR data can be, it is often difficult to distinguish objects such as vehicles in RADAR data alone. As such, in some embodiments, ground truth data may be generated from LIDAR data or other sources of 3D information such as structure from stereo cameras, motion depth estimation, ultrasound, and / or the like. More specifically, a scene may be observed with a RADAR sensor and a LIDAR sensor to collect frames of RADAR data and LIDAR data for a particular time slice. The RADAR data may be used to generate an input RADAR data tensor, and LIDAR data associated with the same or nearest time slice as the RADAR data may be used to generate ground truth labels, which may be used to generate ground truth class distinctions and / or instance regression tensors. More specifically, the LIDAR point cloud may be orthogonally projected to form a LIDAR projection image (e.g., an overhead image) that corresponds to (e.g., has the same size, viewpoint, and / or ground sampling distance) the RADAR projection image contained in the RADAR data tensor. The LIDAR projection image may be annotated (e.g., manually, automatically, etc.) with labels that identify the location, size, orientation, and / or class of instances of associated objects in the LIDAR projection image. The LIDAR labels may comprise bounding boxes, closed polylines, or other bounding shapes drawn, annotated, superimposed, and / or otherwise associated with the LIDAR projection image.

[0019] The LIDAR labels may be used to generate corresponding class confidence tensors and instance regression tensors that may serve as ground truth data for the DNN. In some embodiments, the LIDAR labels may be propagated to the RADAR projection images of the nearest frame of RADAR data (e.g., associated with the same time slice), the number of RADAR detections each LIDAR label contains may be determined, and LIDAR labels that contain fewer than some threshold number of RADAR detections may be omitted. The (remaining) LIDAR labels may be used to generate ground truth data. For example, the position, size, orientation, and / or class of each (remaining) LIDAR label may be used to generate a confidence map that matches the size and dimensionality of the class confidence tensor. As a non-limiting example, for a given class and corresponding dimension of the class confidence tensor, pixel values ​​of pixels that are within the respective labeled bounding shape of that class may be set to a value (e.g., 1) that indicates a positive classification. Additionally or alternatively, the location, size, orientation, and / or class of each (remaining) LIDAR label may be used to generate object information that matches the size and dimensionality of the instance regression tensor. For example, for each pixel that contains a LIDAR label, the LIDAR label may be used to calculate corresponding location, size, and / or orientation information. The orientation information may include, for example, but is not limited to, information related to: where the object is located relative to each pixel (e.g., with respect to the object center), object height, object width, object orientation (e.g., rotation angle relative to the orientation of the projected image), and / or the like. The calculated object information may be stored in a corresponding channel of the instance regression tensor. Thus, the LIDAR labels may be used to generate ground truth class distinctions and / or instance regression tensors.

[0020] As such, the training data may be used to train the DNN to detect moving and stationary obstacles and other objects from RADAR data, and object detection may be provided to an autonomous vehicle driving stack to enable safe planning and control of the autonomous vehicle. Unlike conventional approaches, the present technique may be used to distinguish between stationary obstacles (e.g., cars) and stationary background noise, which is particularly important when navigating cluttered urban environments. Furthermore, embodiments of the present disclosure may provide a simple and effective method for regressing the dimensions and orientation of these obstacles, where conventional methods have struggled or failed entirely. Furthermore, because embodiments of the present disclosure may rely on RADAR data during operation, object detection may be performed in poor weather and at night, conditions where camera-based and LIDAR-based detection techniques degrade and fail.

[0021] Exemplary Object Detection System Referring to FIG. 1, FIG. 1 is a data flow diagram illustrating an example process of an object detection system according to some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are described merely as examples. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, groupings of functions, etc.) may be used in addition to or instead of those illustrated, and some elements may be omitted entirely. Furthermore, many of the elements described herein are functional entities that may be implemented as separate or distributed components or in conjunction with other components, and in any suitable combination and location. Various functions described herein as being performed by an entity may be implemented by hardware, firmware, and / or software. For example, various functions may be implemented by a processor executing instructions stored in a memory.

[0022] At a high level, the process 100 may include one or more machine learning models 108 configured to detect objects, such as obstacle instances, from sensor data 102, such as RADAR detections generated from a RADAR sensor 101. The sensor data 102 may be pre-processed 104 into input data (e.g., RADAR data tensor 106) having a format understood by the machine learning model 108, and the input data may be provided to the machine learning model 108 to detect 116 objects represented in the input data. In some embodiments, the machine learning model 108 predicts class confidence tensors 110 and instance regression tensors 112, which may be post-processed 114 into object detections 116, including bounding boxes, closed polylines, or other bounding shapes that identify the location, size, and / or orientation of the detected objects. These object detections 116 may correspond to obstacles around the autonomous vehicle and may be used by control components of the autonomous vehicle (e.g., controller 1336, ADAS system 1338, SOC 1304, software stack 122, and / or other components of autonomous vehicle 1300 of FIGS. 13A-13D) to assist the autonomous vehicle in performing one or more operations within the environment (e.g., obstacle avoidance, path planning, mapping, etc.).

[0023] In embodiments in which the sensor data 102 includes RADAR data, the RADAR data may be captured with respect to three-dimensional (3D) space. For example, one or more RADAR sensors 101 of an ego object or ego actor (e.g., RADAR sensor 1360 of the autonomous vehicle 1300 of FIGS. 13A-13D) may be used to generate RADAR detection of objects in the environment around the vehicle. In general, a RADAR system may include a transmitter that emits radio waves. The radio waves reflect off certain objects and materials, and one of the RADAR sensors 101 may detect these reflections and their reflection characteristics, such as direction, azimuth, altitude, range (e.g., time of beam flight), brightness, Doppler velocity, RADAR cross section (RCS), reflectivity, SNR, and / or the like. The reflections and reflection characteristics may depend on the objects in the environment, speed, materials, sensor mounting location and orientation, etc. Firmware associated with the RADAR sensor 101 may be used to control the RADAR sensor 101 to capture and / or process sensor data 102, e.g., reflection data from the sensor's field of view. In general, the sensor data 102 may include raw sensor data, RADAR point cloud data, and / or reflection data processed into some other format. For example, the reflection data may be combined with location and orientation data (e.g., from GNSS and IMU sensors) to form a point cloud that represents detected reflections from the environment. Each detection in the point cloud may include metadata about the detection, such as one or more of the detection's three-dimensional location and reflection characteristics.

[0024] The sensor data 102 may be pre-processed 104 into a format understood by the machine learning model 108. For example, in embodiments in which the sensor data 102 includes RADAR detections, the RADAR detections may be accumulated, transformed into a single coordinate system (e.g., centered on the ego actor / vehicle), ego-motion corrected (e.g., to the last known location of the ego actor / vehicle), and / or orthogonally projected to form a projected image (e.g., an overhead image) of a desired size (e.g., spatial dimensions) and having a desired ground sampling distance. The projected image and / or other reflected data may be stored and / or encoded into an appropriate representation, such as a RADAR data tensor 106, which may serve as an input to the machine learning model 108.

[0025] 2 is a data flow diagram illustrating an example process for pre-processing 104 sensor data 102 for a machine learning model 108 in an object detection system, according to some embodiments of the present disclosure. In this example, the sensor data 102 may include RADAR detections, which may be accumulated 210 (which may include transforming to a single coordinate system), ego-motion corrected 220, and / or encoded 230 into a suitable representation, such as a projection image of the RADAR detection, where multiple channels store different reflectance characteristics.

[0026] More specifically, sensor detections, such as RADAR detections, may be accumulated 210 from multiple sensors, such as peripheral RADAR sensors 1360, from different positions of the autonomous vehicle 1300 and transformed to a single vehicle coordinate system (e.g., vehicle-centered). Additionally or alternatively, sensor detections may be accumulated 210 over time to increase the density of accumulated sensor data. Sensor detections may be accumulated over any desired window of time (e.g., 0.5 seconds, 1 second, 2 seconds, etc.). The size of the window may be selected based on the sensor and / or application (e.g., a smaller window may be selected for a noisy application such as a highway scenario). As such, each input to the machine learning model 108 may be generated from accumulated detections from each window of time from a rotating window (e.g., from a duration ranging from t window size to the present). Each window to evaluate may be incremented by any suitable step size, which may not necessarily correspond to a window size. Thus, each successive input to the machine learning model 108 may be based on successive windows, which may not need to overlap.

[0027] In some embodiments, ego-motion correction 220 may be applied to the sensor detections. For example, accumulated detections may be ego-motion corrected to the latest known vehicle location. More specifically, the location of an older detection may be communicated to the latest known location of the moving vehicle using the known motion of the vehicle to estimate where the older detection would be located (e.g., relative to the vehicle's current location) at a desired time (e.g., the current time). The result may be a set of accumulated, ego-motion corrected detections (e.g., RADAR detections) for a particular time slice.

[0028] In some embodiments, the (accumulated, ego-motion corrected) RADAR detections may be encoded 230 into a suitable representation, such as a projected image, which may include multiple channels storing different features, such as reflectance properties. More specifically, the accumulated, ego-motion corrected detections may be orthogonally projected to form a projected image of a desired size (e.g., spatial dimension) and with a desired ground sampling distance. Any desired view of the environment, e.g., top-down, frontal, perspective, and / or other, may be selected for the projected image. In some embodiments, multiple projected images with different views may be generated, with each projected image being input into a separate channel of the machine learning model 108. Because the projected images may be evaluated as inputs to the machine learning model 108, there is generally a trade-off between prediction accuracy and computational needs. As such, the desired spatial dimension and ground sampling distance (e.g., meters per pixel) of the projected image may be selected as a design choice.

[0029] In some embodiments, a projection image may include multiple layers, with pixel values ​​in different layers storing different reflectance characteristics. In some embodiments, for each pixel in a projection image that obtains one or more detections, a set of features may be calculated, determined, or otherwise selected from the reflectance characteristics of the RADAR detections (e.g., bearing, azimuth, altitude, range, brightness, Doppler velocity, RADAR cross section (RCS), reflectivity, signal-to-noise ratio (SNR), etc.). When there are multiple detections obtained on a pixel, thereby forming a tower of points, a particular feature for that pixel may be calculated by aggregating (e.g., using standard deviation, average, etc.) the corresponding reflectance characteristics of the multiple overlapping detections. Thus, any given pixel may have multiple associated feature values ​​that may be stored in corresponding channels of the RADAR data tensor 106. As such, the RADAR data tensor 106 may serve as an input to the machine learning model 108.

[0030] Turning now to FIG. 3, FIG. 3 is a diagram of an example implementation of the machine learning model 108 according to some embodiments of the present disclosure. At a high level, the machine learning model 108 can accept sensor data (e.g., RADAR data processed into the RADAR data tensor 106) as input to detect objects, such as instances of obstacles represented in the sensor data. In a non-limiting example, the machine learning model 108 can receive projection images of accumulated, ego-motion corrected, and orthogonally projected RADAR detections as input, where any given pixel can store various reflectance characteristics of the RADAR detection in a corresponding channel of the input tensor (e.g., the RADAR data tensor 106). To detect objects from the input, the machine learning model 108 can predict classification data (e.g., class confidence tensor 110) and / or object instance data, such as position, size, and / or orientation data for each class (e.g., the instance regression tensor 112). The classification data and object instance data may be post-processed to generate bounding boxes, closed polylines, or other bounding shapes that identify the position, size, and / or orientation of the detected object instances.

[0031] In some embodiments, the machine learning model 108 may be implemented using a DNN, such as a convolutional neural network (CNN). Although certain embodiments are described with the machine learning model 108 being implemented using a neural network, and more specifically, a CNN, this is not intended to be limiting. For example, and without limitation, the machine learning model 108 may include any type of machine learning model, such as linear regression, logistic regression, decision trees, support vector machines (SVMs), naive Bayes, k-nearest neighbors (Knn), K-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, machine learning models using neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid state machines, etc.), and / or other types of machine learning models.

[0032] In general, the machine learning model 108 may include a common trunk (or stream of layers) with several heads (or at least partially discrete streams of layers) for predicting different outputs based on input data. For example, but not limited to, the machine learning model 108 may include a feature extractor including convolutional layers, pooling layers, and / or other layer types, where the output of the feature extractor is provided as input to a first head for predicting classification data and a second head for predicting the position, size, and / or orientation of a detected object. In some instances, the first head and the second head may receive parallel inputs, thereby producing different outputs for similar input data. In the example of FIG. 3, the machine learning model 108 is shown with an exemplary architecture that extracts features from the RADAR data tensor 106 and performs class classification and / or instance regression on the extracted features. More specifically, the machine learning model 108 of FIG. 3 includes a feature extractor trunk 310, a class confidence head 320, and an instance regression head 330.

[0033] The feature extractor trunk 310 may be implemented using encoder and decoder components with skip connections (e.g., similar to a feature pyramid network, U-Net, etc.). For example, the feature extractor trunk 310 may accept input data, such as the RADAR data tensor 106, and apply various convolution, pooling, and / or other types of operations to extract features in some latent space. In FIG. 3, the feature extractor trunk 310 is shown with an example implementation including an encoder / decoder with an encoding (contracting) path at the bottom left and an example decoding (expanding) path at the top right. Along the contracting path, each decomposition may include any number of layers (e.g., convolution, dilated convolution, inception block, etc.) and downsampling operations (e.g., max pooling). Along the expanding path, each decomposition may include any number of layers (e.g., deconvolution, upsampling followed by convolution, and / or other types of operations). In the expanding path, each decomposition of the feature map may be upsampled and concatenated (e.g., in the depth dimension) with the feature map of the same decomposition from the contracting path. In this example, corresponding decompositions of the contracting and expanding paths may be connected with skip connections (e.g., skip connections 312) that can be used to append or concatenate feature maps from corresponding decompositions (e.g., forming concatenated feature map 314). As such, the feature extractor trunk 310 can extract features into several latent space tensors that may be input to the class confidence head 320 and the instance regression head 330.

[0034] The class confidence head 320 may include any number of layers 320A, 320B, 320C (e.g., convolution, pooling, softmax, and / or other types of classifiers, etc.) that predict classification data from the output of the feature extractor trunk 310. For example, the class confidence head 320 may include a channel (e.g., a stream of classifiers in layers) for each class of object to be detected (e.g., vehicles, cars, trucks, vulnerable road users, pedestrians, cyclists, motorcycles, etc.), such that the class confidence head 320 is responsible for predicting classification data (e.g., confidence maps) in the form of a multi-channel tensor (e.g., class confidence tensor 110). Each channel can be thought of as a heat map with a confidence / likelihood value for each pixel belonging to the class corresponding to the channel.

[0035] The instance regression head 330 may include any number of layers 330A, 330B, 330C (e.g., classifiers such as convolution, pooling, softmax, and / or other types of operations) that predict object instance data (e.g., location, size, and / or orientation of detected objects) from the output of the feature extractor trunk 310. The instance regression head 330 may include N channels (e.g., a stream of classifiers plus layers), where each channel regresses a particular type of information about detected object instances of that class, such as where the object is located (e.g., a dx / dy vector pointing to the center of the object), the height of the object, the width of the object, the orientation of the object (e.g., rotation angles such as sine and / or cosine), and / or the like. As a non-limiting example, the instance regression head 330 may include separate dimensions identifying the x dimension of the center of the detected object, the y dimension of the center of the detected object, the width of the detected object, the height of the detected object, the sine of the orientation of the detected object (e.g., the rotation angle in 2D image space), the cosine of the orientation of the detected object, and / or other types of information. These types of object instance data are intended as examples only, and other types of object information may be regressed within the scope of this disclosure. Thus, the instance regression head 330 may be responsible for predicting a multi-channel instance regression tensor (e.g., the instance regression tensor 112) that stores N types of object information. Each channel of the instance regression tensor 112 may include a floating point number that regresses a particular type of object information, such as a particular object dimension.

[0036] As such, the machine learning model 108 can predict multi-channel classification data (e.g., class confidence tensor 110) and / or multi-channel object instance data (e.g., instance regression tensor 112) from a particular input (e.g., RADAR data tensor 106). Some possible training techniques are described in more detail below. In operation, the output of the machine learning model 108 can be post-processed (e.g., decoded) to generate bounding boxes, closed polylines, or other bounding shapes that identify the location, size, and / or orientation of the detected object instances, as described in more detail below. In addition to or as an alternative to the machine learning model 108 using a common trunk with separate partitioning heads, separate DNN featurizers can be configured to evaluate projection images from different views of the environment. In one example, multiple projection images may be generated at different views, each projection image may be fed into a separate side-by-size DNN featurizer, and the latent space tensors of the DNN featurizers may be combined and decoded into object detection (e.g., bounding boxes, closed polylines, or other bounding shapes). In another example, sequential DNN featurizers may be chained. In this example, a first projection image may be generated at a first view of the environment (e.g., a perspective view), the first projection image may be fed into a first DNN featurizer (e.g., predicting classification data), and the output of the first DNN featurizer may be transformed to a second view of the environment (e.g., a top-down view), which may be fed into a second DNN featurizer (e.g., predicting object instance data). These architectures are intended as merely illustrative, and other architectures (single-view or multi-view scenarios with separate DNN featurizers) are contemplated within the scope of this disclosure.

[0037] As mentioned above, the output of the machine learning model 108 may be post-processed (e.g., decoded) to generate a bounding box, closed polyline, or other bounding shape that identifies the location, size, and / or orientation of the detected object instance. For example, when the input to the machine learning model 108 includes a projection image (e.g., of the accumulated, ego-motion corrected, and orthogonally projected RADAR detections), a bounding box, closed polyline, or other bounding shape may be identified for the projection image (e.g., in the image space of the projection image). In some embodiments, because the object instance data may be noisy and / or may yield multiple candidates, the bounding shape may be generated using non-maximum suppression, density-based spatial clustering of application with noise (DBSCAN), and / or another function.

[0038] FIG. 4 is a data flow diagram illustrating an example post-processing process 114 for generating object detections 116 in an object detection system according to some embodiments of the present disclosure. In this example, the post-processing process 114 includes an instance decoder 140 and filtering and / or clustering 420. In general, the instance decoder 140 can identify candidate bounding boxes (or other bounding shapes) (e.g., for each object class) based on the object instance data (e.g., position, size, and / or orientation data) from a corresponding channel of the instance regression tensor 112 and / or a confidence map from a corresponding channel of the class confidence tensor 110 for that class. More specifically, the predicted confidence map and the predicted object instance data can specify information about the detected object instance, such as where the object is located, object height, object width, object orientation, and / or the like. This information can be used to identify candidate object detections (e.g., candidates with unique center points, object height, object width, object orientation, and / or the like). The result may be a set of candidate bounding boxes (or other bounding shapes) for each object class.

[0039] Various types of filtering and / or clustering 420 may be applied to remove duplicates and / or noise from the candidate bounding boxes (or other bounding shapes) for each object class. For example, in some embodiments, duplicates may be removed using non-maximum suppression. Non-maximum suppression may be used when two or more candidate bounding boxes have associated confidence values ​​that indicate the candidate bounding boxes may correspond to the same object instance. In such instances, the highest confidence value of an object instance may be used to determine which candidate bounding box should be used for that object instance, and non-maximum suppression may be used to remove or suppress other candidates.

[0040] For example, each candidate bounding box (or other bounding shape) may be associated with a corresponding confidence / likelihood value associated with one or more corresponding pixels from a corresponding channel of the class confidence tensor 110 of the class being evaluated (e.g., using the confidence / likelihood value of a representative pixel such as a central pixel, using an average or some other composite value calculated over the candidate region, etc.). Thus, candidate bounding shapes with confidence / likelihood of being a member of the object class below some threshold (e.g., 50%) may be filtered out. Additionally or alternatively, the candidate bounding box (or other shape) with the highest confidence / likelihood score for a particular class may be assigned an instance ID, a metric such as intersection of unions (IoU) may be calculated for each of the other candidates in the class, and candidates with IoU above some threshold may be filtered to remove duplicates. The process may then be repeated, assigning an instance ID to the candidate with the highest confidence / likelihood score, removing duplicates, and repeating until no candidates remain. The process may be repeated for each of the other classes to remove duplicate candidates.

[0041] In some embodiments, a clustering technique such as Density-Based Spatial Clustering with Noise (DBSCAN) may be used to remove duplicate candidate boundary shapes. For example, the candidate boundary shapes may be clustered (e.g., the centers of the candidate boundary shapes may be clustered), candidates within each cluster may be determined to correspond to the same object instance, and duplicate candidates from each cluster may be removed.

[0042] In summary, the machine learning model 108 can accept sensor data such as projected images (e.g., of accumulated, ego-motion corrected, and orthogonally projected RADAR detections) and predict classification data and / or object instance data that can be post-processed to generate bounding boxes, closed polylines, or other bounding shapes that identify the location, size, and / or orientation of detected object instances in the projected images. Figure 5A is a diagram of example orthogonal projections (i.e., white bounding boxes in this example) of accumulated RADAR detections and corresponding object detections according to some embodiments of the present disclosure. For visualization purposes, Figure 5B is a diagram of object detections projected onto corresponding images from three cameras.

[0043] Once the position, size, and / or orientation of the object instance have been determined, the 2D pixel coordinates defining the object instance may be transformed into 3D world coordinates for use by the autonomous vehicle in performing one or more operations (e.g., obstacle avoidance, lane keeping, lane changing, path planning, mapping, etc.). More specifically, returning to FIG. 1, the object detection 116 (e.g., bounding box, closed polyline, or other bounding shape) may be used by a control component of the autonomous vehicle 1300 shown in FIGS. 13A-13D, such as an autonomous driving software stack 122 executing on one or more components (e.g., SoC 1304, CPU 1318, GPU 1320, etc.) of the vehicle 1300. For example, the vehicle 1300 may use this information (e.g., obstacle instances) to navigate, plan, or otherwise perform one or more operations (e.g., obstacle avoidance, lane keeping, lane changing, merging, splitting, etc.) within the environment.

[0044] In some examples, object detection 116 may be used by one or more layers of an autonomous driving software stack 122 (alternatively referred to herein as “driving stack 122”). Driving stack 122 may include a sensor manager (not shown), a perception component (e.g., corresponding to a perception layer of driving stack 122), a world model manager 126, a planning component 128 (e.g., corresponding to a planning layer of driving stack 122), a control component 130 (e.g., corresponding to a control layer of driving stack 122), an obstacle avoidance component 132 (e.g., corresponding to an obstacle or collision avoidance layer of driving stack 122), an actuation component 134 (e.g., corresponding to an actuation layer of driving stack 122), and / or other components corresponding to additional and / or alternative layers of driving stack 122. In some instances, process 100 may be performed by the perception component, which may provide the layers of driving stack 122 to the world model manager, as described in more detail herein.

[0045] The sensor manager can manage and / or abstract sensor data 102 from sensors of the vehicle 1300. For example, referring to FIG. 13C, the sensor data 102 can be generated by a RADAR sensor 1360 (e.g., permanently, at intervals, based on certain conditions). The sensor manager can receive sensor data 102 from sensors in different formats (e.g., sensors of the same type can output sensor data in different formats) and can be configured to convert the different formats to a unified format (e.g., for each sensor of the same type). As a result, other components, features, and / or functionality of the autonomous vehicle 1300 can use the unified format, thereby simplifying the processing of the sensor data 102. In some instances, the sensor manager can use the unified format to apply control back to sensors of the vehicle 1300, for example, to set frame rate or to perform gain control. The sensor manager may also update sensor packets or communications corresponding to the sensor data with a timestamp to help inform the processing of the sensor data by the various components, features, and functionality of the autonomous vehicle control system.

[0046] The world model manager 126 may be used to generate, update, and / or define a world model. The world model manager 126 may use information generated by and received from the perception components (e.g., location of detected obstacles) of the driving stack 122. The perception components may include obstacle perception devices, path perception devices, wait perception devices, map perception devices, and / or other perception components. For example, the world model may be defined, at least in part, based on the affordances of obstacles, paths, and wait states that may be perceived in real-time or near real-time by the obstacle perception devices, path perception devices, wait perception devices, and / or map perception devices. The world model manager 126 may continuously update the world model based on newly generated and / or received input (e.g., data) from the obstacle perception devices, path perception devices, wait perception devices, map perception devices, and / or other components of the autonomous vehicle control system.

[0047] The world model can be used to help inform the planning component 128, the control component 130, the obstacle avoidance component 132, and / or the actuation component 134 of the driving stack 122. The obstacle perception device can perform obstacle perception, which can be based on where the vehicle 1300 is allowed or able to drive (e.g., based on the location of a drivable path defined by avoiding detected obstacles) and how fast the vehicle 1300 can drive without colliding with an obstacle (e.g., an object such as a structure, entity, vehicle, etc.) detected by sensors of the vehicle 1300 and / or the machine learning model 108.

[0048] The path perception device can perform path perception, such as by perceiving the nominal paths available in a particular situation. In some instances, the path perception device can further consider lane changes for path perception. The lane graph can represent one or more paths available to the vehicle 1300 and can be as simple as a single path on a highway on-ramp. In some instances, the lane graph can include paths to a desired lane and / or indicate available changes onto a highway (or other road type), or can include nearby lanes, lane changes, forks, turns, cloverleaf interchanges, merges, and / or other information.

[0049] The wait perception device can be responsible for determining constraints of the vehicle 1300 as a result of rules, conventions, and / or practical considerations. For example, the rules, conventions, and / or practical considerations may relate to traffic signals, multi-way stops, yield ahead, merging, toll booths, gates, police or other emergency personnel, road workers, stopped buses or other vehicles, one-way bridge arbitration, ferry entrances, etc. Thus, the wait perception device can be utilized to identify potential obstacles and to implement one or more controls (e.g., slowing down, stopping, etc.) that may not be possible by relying solely on the obstacle perception device.

[0050] The map perception device may include mechanisms by which movements are discriminated and, in some instances, for determining the particular instance of which conventions apply to a particular locale. For example, the map perception device may determine from data representing a previous drive or journey that a particular intersection does not allow U-turns at a particular time of day, that the electronic signs indicating lane directionality change depending on the time of day, that two traffic signals in close proximity (e.g., slightly offset from one another) are associated with different roads, that in Rhode Island, a first passenger vehicle waiting to turn left at a traffic light will violate the law by turning in front of oncoming traffic when the light turns green, and / or other information. The map perception device may inform the vehicle 1300 of static or stationary infrastructure objects and obstacles. The map perception device may also generate information for the wait perception device and / or the path perception device, such as to determine which lights at an intersection must be green for the vehicle 1300 to take a particular path.

[0051] In some instances, information from the map perception device can be sent, transmitted, and / or provided to a server (e.g., to a map manager of server 1378 of FIG. 13D ), and information from the server can be sent, transmitted, and / or provided to the map perception device and / or localization manager of vehicle 1300. The map manager can include a cloud mapping application located remotely from vehicle 1300 and accessible by vehicle 1300 via one or more networks. For example, the map perception device and / or localization manager of vehicle 1300 can communicate with the map manager of the server and / or one or more other components or functions to inform the map perception device and / or localization manager of the past and current driving or travel of vehicle 1300, as well as the past and current driving or travel of other vehicles. The map manager can provide mapping output (e.g., map data) that can be localized by the localization manager based on the particular location of the vehicle 1300, and the localized mapping output can be used by the world model manager 126 to generate and / or update a world model.

[0052] The planning component 128 may include a route planner, a lane planner, a motion planner, and a motion selector, among other components, features, and / or functionalities. The route planner may use information from the map perception device, the map manager, and / or the localization manager, among other information, to generate a planned path, which may consist of GNSS waypoints (e.g., GPS waypoints), 3D world coordinates (e.g., Cartesian, polar, etc.) indicating coordinates relative to an origin of the vehicle 1300, and the like. The waypoints may represent a particular distance into the future of the vehicle 1300, e.g., a number of city blocks, a number of kilometers, a number of feet, a number of inches, a number of miles, etc., that may be used as a target for the lane planner.

[0053] The lane planner can use as input a target point and direction in a lane graph (e.g., a lane graph from a path perception device), object poses in the lane graph (e.g., from a localization manager), and / or a distance into the future from a route planner. The target point and direction can be mapped to a best matching drivable point and direction in the lane graph (e.g., based on GNSS and / or compass orientation). A graph search algorithm can be run on the lane graph from the current edge in the lane graph to find the shortest path to the target point.

[0054] The motion planner can determine the feasibility of basic motions of the vehicle 1300, such as staying in the lane or changing lanes to the left or right, so that feasible motions can be matched with the most desirable motions output from the lane planner. For example, if a desired motion is determined to be unsafe and / or available, a default motion can be selected instead (e.g., when a desired motion or lane change is unsafe, the default motion can be to stay in the lane).

[0055] The control component 130 can follow the trajectory or path (lateral and longitudinal) received from the motion selector of the planning component 128 (e.g., based on the object detection 116) as closely as possible and within the capabilities of the vehicle 1300. The control component 130 can use hard feedback to address unplanned events or movements that were not modeled and / or anything that causes a deviation from the ideal (e.g., unexpected delays). In some instances, the control component 130 can use a forward prediction model that receives the controls as input variables and produces a prediction that can be compared to the desired state (e.g., compared to the desired lateral and longitudinal paths requested by the planning component 128). Controls that minimize the deviation can be determined.

[0056] Although the planning component 128 and the control component 130 are shown separately, this is not intended to be limiting. For example, in some embodiments, the delineation between the planning component 128 and the control component 130 may not be precisely defined. As such, at least some of the components, features, and / or functionality attributed to the planning component 128 may be associated with the control component 130, and vice versa. This may also be true for any of the separately shown components of the driving stack 122.

[0057] The obstacle avoidance component 132 can assist the autonomous vehicle 1300 in avoiding collisions with objects (e.g., moving and stationary objects). The obstacle avoidance component 132 can include computational mechanisms at the “root level” of obstacle avoidance and can act as the “survival brain” or “reptilian brain” of the vehicle 1300. In some instances, the obstacle avoidance component 132 can be used independent of the components, features, and / or functionality of the vehicle 1300 that are required to obey traffic rules and drive courteously. In such instances, the obstacle avoidance component can ignore traffic laws, rules of the road, and courteous driving norms to ensure that no collisions occur between the vehicle 1300 and objects. As such, the obstacle avoidance layer can be a separate layer from the road layer rules, which can ensure that the vehicle 1300 is only performing safe actions from an obstacle avoidance perspective. On the other hand, the road layer rules can ensure that the vehicle obeys traffic laws and customs and adheres to legal conventional right-of-way (as described above).

[0058] In some instances, the drivable path and / or object detection 116 may be used by the obstacle avoidance component 132 in determining control or action to take. For example, the drivable path may provide instructions to the obstacle avoidance component 132 of where the vehicle 1300 may maneuver without colliding with objects, structures, and / or the like, or at least where static structures may not be present.

[0059] In non-limiting examples, obstacle avoidance component 132 may be implemented as a separate, discrete feature of vehicle 1300. For example, obstacle avoidance component 132 may operate separately from (e.g., in parallel with, before, and / or after) the planning layer, the control layer, the actuation layer, and / or other layers of driving stack 122.

[0060] As such, the vehicle 1300 can use this information (e.g., as a road edge, or a barrier) to navigate, plan, or otherwise perform one or more actions within the environment (e.g., lane keeping, lane changing, merging, splitting, etc.).

[0061] 6-8, each block of the methods 600, 700, and 800 described herein includes computational processes that may be performed using any combination of hardware, firmware, and / or software. For example, various functions may be performed by a processor executing instructions stored in a memory. The methods may also be implemented as computer usable instructions stored on a computer storage medium. The methods may be provided by a standalone 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. Additionally, the methods 600, 700, and 800 are described, by way of example, with respect to the object detection systems described herein. However, the methods may additionally or alternatively be performed by any one system or any combination of systems, including but not limited to those described herein.

[0062] 6 is a flow diagram illustrating a method 600 of object detection using RADAR data according to some embodiments of the disclosure. The method 600 includes, at block B602, accumulating RADAR data representing sensor detections of one or more RADAR sensors of an ego actor, such as a moving vehicle, over a period of time. For example, the RADAR detections may be accumulated from multiple sensors, e.g., some or all of the peripheral RADAR sensors 1360 from different positions of the autonomous vehicle 1300, and transformed to a single vehicle coordinate system (e.g., vehicle-centered). Additionally or alternatively, the sensor detections may be accumulated over time to increase the density of the accumulated sensor data.

[0063] The method 600 includes generating motion-compensated sensor detections at block B604. For example, accumulated detections may be ego-motion compensated to the last known location of an ego actor (e.g., a moving vehicle). The location of an old detection may be communicated to the last known location of the moving vehicle using the known motion of the vehicle to estimate where the old detection would currently be located (e.g., relative to the vehicle's current position).

[0064] The method 600 includes projecting the motion-compensated sensor detections into a two-dimensional (2D) image space to generate a projected image at block B606. For example, the motion-compensated sensor detections may be orthogonally projected to form a projected image (e.g., from a top-down view) of a desired size (e.g., spatial dimensions) and having a desired ground sampling distance.

[0065] The method 600 includes applying the data representing the projection images to a neural network at block B608. For example, the projection images and corresponding reflectance characteristics can be stored in multiple layers of tensors, with pixel values ​​in different layers storing different reflectance characteristics. In some embodiments, for each pixel of the projection images that yields one or more detections, a set of features can be calculated, determined, or selected from the reflectance characteristics of the RADAR detections (e.g., bearing, azimuth, altitude, range, brightness, Doppler velocity, RADAR cross section (RCS), reflectivity, signal-to-noise ratio (SNR), etc.).

[0066] The method 600 includes, at block B610, using a neural network to calculate outputs representative of one or more detected objects based at least in part on the data representative of the projection images. For example, the neural network can accept the projection images and corresponding reflectance properties encoded in a suitable representation, such as a tensor, and predict outputs such as classification data (e.g., a class confidence tensor) and / or object instance data for each detected object instance (e.g., an instance regression tensor).

[0067] The method 600 includes generating at least one bounding shape corresponding to each of the one or more detected objects based at least in part on the output of the neural network at block B612. For example, the output may include classification data and / or object instance data that may be used to calculate a number of candidate bounding shapes, and the candidate bounding shapes may be filtered and / or clustered to remove duplicates.

[0068] 7 is a flow diagram illustrating a method 700 for object detection using RADAR data, according to some embodiments of the present disclosure. The method 700 includes, at block B702, accessing a multi-channel RADAR data tensor that encodes accumulated ego-motion corrected sensor detections and corresponding reflectance characteristics corresponding to an environment of an ego actor. For example, the multi-channel RADAR data tensor may include a projection image of the accumulated ego-motion corrected sensor detections with pixel values ​​of different channels of the multi-channel RADAR data tensor that store different reflectance characteristics.

[0069] The method 700 includes applying the data representing the multi-channel RADAR data tensor to a neural network at block B704 to cause the neural network to compute an output indicative of one or more detected objects in the environment. For example, the neural network may accept the multi-channel RADAR data tensor, corresponding projection images, and / or corresponding reflectance properties and predict outputs such as classification data (e.g., a class confidence tensor) and / or object instance data (e.g., an instance regression tensor) for each detected object instance.

[0070] The method 700 includes generating one or more bounding shapes from the output of the neural network at block B706 that correspond to one or more detected objects in the environment. For example, the output may include classification data and / or object instance data that may be used to compute a number of candidate bounding shapes, which may be filtered and / or clustered to remove duplicates.

[0071] 8 is a flow diagram illustrating a method 800 of object detection using RADAR data according to some embodiments of the disclosure. The method 800 includes, at block B802, accumulating sensor data representing sensor detections and corresponding reflective characteristics corresponding to an environment of an ego actor, e.g., a moving vehicle, over a period of time. For example, the RADAR detections may be accumulated from multiple sensors, e.g., some or all of the peripheral RADAR sensors 1360 from different locations of the autonomous vehicle 1300, and transformed to a single vehicle coordinate system (e.g., vehicle-centered). Additionally or alternatively, the sensor detections may be accumulated over time to increase the density of the accumulated sensor data.

[0072] The method 800 includes, in block B804, converting the sensor detection to a latest known location of the ego actor to generate a motion-corrected sensor detection. For example, the sensor detection may be ego-motion corrected to the latest known location of the ego actor (e.g., a moving vehicle). The location of an old detection may be communicated to the latest known location of the moving vehicle using the known motion of the vehicle to estimate where the old detection would currently be located (e.g., relative to the vehicle's current position).

[0073] The method 800 includes projecting the motion-compensated sensor detections into a projection image at block B806. For example, the motion-compensated sensor detections may be orthogonally projected to form a projection image (e.g., from a top-down view) of a desired size (e.g., spatial dimensions) and having a desired ground sampling distance.

[0074] The method 800 includes, at block B808, calculating, for each pixel of the projection image, a set of features based at least in part on reflectance characteristics corresponding to a subset of the motion-compensated sensor detections associated with that pixel. For example, in some embodiments, for each pixel of the projection image that obtains one or more detections, a set of features may be calculated, determined, or selected from the reflectance characteristics of the RADAR detections (e.g., bearing, azimuth, altitude, range, brightness, Doppler velocity, RADAR cross section (RCS), reflectivity, signal-to-noise ratio (SNR), etc.). When there are multiple detections obtained at a pixel, thereby forming a tower of points, a particular feature may be calculated by aggregating (e.g., using standard deviation, average, etc.) the corresponding reflectance characteristics of the multiple overlapping detections.

[0075] The method 800 includes generating an output encoding the projection images and a set of features of a neural network configured to detect objects in the environment at block B810. For example, the projection images and corresponding reflectance properties may be stored in multiple layers of tensors, with pixel values ​​in different layers storing different reflectance properties.

[0076] Training Machine Learning Models for Object Detection Systems To train a machine learning model (e.g., machine learning model 108 of FIG. 1 ) of an object detection system, training data may be generated from the sensor data using the preprocessing techniques described above (e.g., preprocessing 104 of FIG. 1 ). However, given how sparse the RADAR data can be, it is often difficult to distinguish objects, such as vehicles, in the RADAR data alone. As such, in some embodiments, ground truth data may be generated from LIDAR data generated from one or more LIDAR sensors.

[0077] For example, a scene may be observed with RADAR and LIDAR sensors (e.g., RADAR sensor 1360 and LIDAR sensor 1364 of autonomous vehicle 1300 of FIGS. 13A-13D) to collect frames of RADAR and LIDAR data for a particular time slice. The RADAR data may be used to generate an input RADAR data tensor using the preprocessing techniques described above (e.g., preprocessing 104 of FIG. 1). LIDAR data associated with the same or nearest time slice as the RADAR data may be used to generate ground truth labels, which may be used to generate ground truth classification data and / or object instance data corresponding to the output of machine learning model 108 (e.g., class confidence tensor 110 and instance regression tensor 112 of FIG. 1).

[0078] More specifically, the LIDAR point cloud may be orthogonally projected to form a LIDAR projection image (e.g., an overhead image) that corresponds to a RADAR projection image contained in the RADAR data tensor (e.g., having the same dimensionality, viewpoint, and / or ground sampling distance). The LIDAR projection image may be annotated (e.g., manually, automatically, etc.) with labels that identify the location, size, orientation, and / or class of instances of associated objects in the LIDAR projection image. The LIDAR labels may be generated within a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of suitable program, and / or may be hand-drawn in some instances. In any instance, LIDAR labels can be synthetically produced (e.g., generated from a computer model or rendering), realistically produced (e.g., designed and produced from real-world data), automated by a machine (e.g., using feature analysis to extract features from data and then learning to generate labels), human-annotated (e.g., a labeler or annotation expert defines the location of the label), and / or a combination thereof (e.g., a human identifies polyline intersections and a machine generates polygons using a polygon rasterizer). In general, LIDAR labels can include bounding boxes, closed polylines, or other bounding shapes that are drawn, annotated, superimposed, and / or otherwise associated with the LIDAR projection image. FIG. 9B is a diagram of an example LIDAR frame with a white ground truth bounding box.

[0079] In general, LIDAR labels identified from LIDAR frames may be used to generate corresponding class confidence tensors and instance regression tensors that may serve as ground truth data for the DNN. However, certain LIDAR labels may be selected for use in generating ground truth data based on corresponding RADAR data. More specifically, LIDAR labels identified from LIDAR frames may be propagated to corresponding RADAR projection images of nearest frames of RADAR data (e.g., associated with the same time slice). In some embodiments, the LIDAR projection images and the RADAR projection images may correspond to one another (e.g., have the same dimensionality, viewpoint, and / or ground sampling distance). This may facilitate direct comparison of LIDAR detections to RADAR detections and evaluation of which LIDAR labels correspond to observed RADAR detections. For example, the number of RADAR detections each propagated LIDAR label contains may be determined, and LIDAR labels that contain fewer RADAR detections than some threshold number may be omitted from use in generating ground truth data. As such, the remaining LIDAR labels may be used as ground truth labels to generate ground truth data for the machine learning model 108.

[0080] FIG. 9A is a diagram of an example orthogonal projection of accumulated RADAR detections and ground truth bounding boxes (shown in white) propagated from a corresponding LIDAR frame, and FIG. 9B is a diagram of a corresponding LIDAR frame with ground truth bounding boxes, according to some embodiments of the present disclosure. Note that the LIDAR detections in FIG. 9B are denser than the RADAR detections in FIG. 9A. As such, denser LIDAR detections may be easier to label than sparser RADAR detections. As shown in FIGS. 9A and 9B, the LIDAR labels (shown in white) from the LIDAR frame (more specifically, the LIDAR projection image of the LIDAR frame) shown in FIG. 9B have been propagated to the corresponding RADAR frame (more specifically, the RADAR projection image of the RADAR frame) shown in FIG. 9A. In FIG. 9A, propagated LIDAR labels that include fewer than some threshold number of RADAR detections have been omitted. The remaining LIDAR labels may be used as ground truth labels to generate ground truth data for the machine learning model 108.

[0081] To generate ground truth data from the (remaining) LIDAR labels, the position, size, orientation, and / or class of each (remaining) LIDAR label may be used to generate a confidence map that matches the size and dimensionality of the class confidence tensor 110. As a non-limiting example, for a given class and corresponding dimension of the class confidence tensor 110, the pixel values ​​of pixels contained within each labeled bounding shape of that class may be set to a value indicating a positive classification (e.g., 1). The process may be repeated and stored in the corresponding channel of the ground truth class confidence tensor. Additionally or alternatively, the position, size, orientation, and / or class of each (remaining) LIDAR label may be used to generate object instance data that matches the size and dimensionality of the instance regression tensor 112. For example, for each pixel that contains a LIDAR label, the LIDAR label may be used to calculate corresponding location, size, and / or orientation information (e.g., where the object is located for each pixel (e.g., the center of the object), the object height, the object width, the object orientation (e.g., the rotation angle relative to the orientation of the projected image), and / or the like). The calculated object instance data may be stored in corresponding channels of a ground truth instance regression tensor. Thus, the LIDAR labels may be used to generate ground truth class distinctions and / or instance regression tensors, which may be used with the corresponding RADAR data tensors as part of a training data set to train the machine learning model 108.

[0082] More specifically, a training set of RADAR data tensors may be input to the machine learning model 108, and the ground truth class partitions and / or instance regression tensors may be used to train the machine learning model 108. For example, one or more loss functions (e.g., a single loss function, a loss function for each output type, etc.) may be used to compare the accuracy of the output of the machine learning model 108 to the ground truth, and the parameters of the machine learning model 108 may be updated (e.g., using backward passes, back propagation, forward passes, etc.) until the accuracy reaches an optimal or acceptable level. In some embodiments where the machine learning model 108 includes multiple heads, the multiple heads may be co-trained with a common trunk or may be trained individually. In embodiments including a multi-view scenario with separate DNN featurizers, the DNN featurizers may be co-trained together or separately. These and other variations are contemplated within this disclosure.

[0083] 10-12, each block of the methods 1000, 1100, and 1200 described herein includes computational processes that may be performed using any combination of hardware, firmware, and / or software. For example, various functions may be performed by a processor executing instructions stored in a memory. The methods may also be implemented as computer usable instructions stored on a computer storage medium. The methods may be provided by a standalone 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. Additionally, the methods 900, 1000, and 1100 are described with respect to the training process of the object detection system described above, as an example. However, the methods may additionally or alternatively be performed by any one system or any combination of systems, including but not limited to those described herein.

[0084] 10 is a flow diagram illustrating a method 1000 for training a machine learning model for object detection according to some embodiments of the present disclosure. The method 1000 includes receiving RADAR and LIDAR data collected from a three-dimensional (3D) environment at block B1002. For example, the scene may be observed with RADAR and LIDAR sensors (e.g., RADAR sensor 1360 and LIDAR sensor 1364 of autonomous vehicle 1300 of FIGS. 13A-13D) to collect the RADAR and LIDAR data.

[0085] The method 1000 includes, at block B1004, orthogonally projecting a first instance of the LIDAR data associated with the time slice to generate a LIDAR projection image representative of the LIDAR detection within the 3D environment.

[0086] Method 1000 includes receiving a LIDAR label identifying a location of an object represented by the LIDAR detection in the LIDAR projection image at block B1006. For example, the LIDAR label may be generated within a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of suitable program, and / or may be drawn by hand.

[0087] The method 1000 includes orthogonally projecting the second instance of the RADAR data associated with the time slice to generate a RADAR projection image representative of the RADAR detection within the 3D environment, at block B1008. The LIDAR projection image and the RADAR projection image may correspond to one another (e.g., have the same dimensionality, viewpoint, and / or ground sampling distance).

[0088] The method 1000 includes, at block B1010, imparting the LIDAR label to the RADAR projection image to generate a imparted LIDAR label.

[0089] The method 1000 includes, at block B1012, determining a number of RADAR detections corresponding to each transmitted LIDAR label among the transmitted LIDAR labels.

[0090] The method 1000 includes, at block B1014, removing a set of transmitted LIDAR labels that include fewer than a threshold number of RADAR detections, leaving a remaining set of LIDAR labels.

[0091] The method 1000 includes training a neural network to detect objects from the RADAR projection images using ground truth data generated based at least in part on the remaining set of LIDAR labels at block B1016. For example, ground truth data that matches the size and dimensionality of the output of the neural network (e.g., machine learning model 108) may be generated using the position, size, orientation, and / or class of each LIDAR label from the remaining set of LIDAR labels, and the neural network may be trained using the ground truth data.

[0092] 11 is a flow diagram illustrating a method 1100 for training a machine learning model for object detection according to some embodiments of the present disclosure. The method 1100 includes receiving RADAR and LIDAR data collected from a three-dimensional (3D) environment at block B1102. For example, the scene may be observed with RADAR and LIDAR sensors (e.g., RADAR sensor 1360 and LIDAR sensor 1364 of autonomous vehicle 1300 of FIGS. 13A-13D) to collect the RADAR and LIDAR data.

[0093] Method 1100 includes, in block B1104, performing the steps illustrated in blocks B1104-B1110 for each frame of LIDAR data. For example, in block B1104, method 1100 includes, for each frame of LIDAR data, generating a LIDAR projection image from the frame of LIDAR data that represents LIDAR detections from the 3D environment. For example, LIDAR detections (e.g., LIDAR point clouds) from the frame of LIDAR data may be orthogonally projected to form a LIDAR projection image (e.g., an overhead image).

[0094] The method 1100 includes receiving, for each frame of LIDAR data, a LIDAR label that identifies a location of an object represented by the LIDAR detection in the LIDAR projection image, at block B1106. For example, the LIDAR label may be generated within a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of suitable program, and / or may be drawn by hand.

[0095] The method 1100 includes, at block B1108, generating, for each frame of LIDAR data, a RADAR projection image representative of RADAR detections from the 3D environment from a corresponding frame of RADAR data associated with the frame of LIDAR data. For example, the RADAR detections (e.g., a RADAR point cloud) from the corresponding frame of RADAR data may be orthogonally projected to form a RADAR projection image (e.g., an overhead image).

[0096] The method 1100 includes, for each frame of LIDAR data, training a neural network to detect objects from the RADAR projection images using ground truth data generated based at least in part on the LIDAR labels, at block B1110. For example, ground truth data that matches the size and dimensionality of the output of the neural network (e.g., machine learning model 108) may be generated using the location, size, orientation, and / or class of the LIDAR labels, and the neural network may be trained using the ground truth data.

[0097] 12 is a flow diagram illustrating a method 1200 for conveying LIDAR labels to RADAR data according to some embodiments of the present disclosure. The method 1200 includes, at block B1202, receiving a LIDAR label of a LIDAR detection in one or more instances of LIDAR data collected from an environment over a period of time using a LIDAR sensor. For example, a scene may be observed with a LIDAR sensor (e.g., LIDAR sensor 1364 of autonomous vehicle 1300 of FIGS. 13A-13D) to collect the LIDAR data. The LIDAR label may be generated within a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of suitable program, and / or may be hand-drawn.

[0098] Method 1200 includes, in block B1204, performing the steps illustrated in blocks B1204-B1206 for each of the one or more instances of LIDAR data. For example, in block B1204, method 1200 includes identifying, for each of the one or more instances of LIDAR data, a closest instance of RADAR data to the instance of LIDAR data that is associated with an accumulated, ego-motion corrected, orthogonal image of a portion of the RADAR data collected from the environment over the period of time using the RADAR sensor, the closest instance of RADAR data to the instance of LIDAR data. For example, a scene may be observed with both a RADAR sensor and a LIDAR sensor (e.g., RADAR sensor 1360 and LIDAR sensor 1364 of autonomous vehicle 1300 of FIGS. 13A-13D) to collect RADAR data and LIDAR data, and the closest instance of RADAR data to an instance of LIDAR data may be identified based on the RADAR data and LIDAR data being associated with the same time slice.

[0099] The method 1200 includes, for each of one or more instances of the LIDAR data, communicating a set of LIDAR labels of the LIDAR detections in the instance of the LIDAR data to a corresponding set of RADAR detections in a nearest instance of the RADAR data, at block B1206. A determination may be made of how many of the RADAR detections are included in each of the communicated LIDAR labels, and communicated LIDAR labels that include fewer than some threshold number of RADAR detections (e.g., 2, 5, etc.) may be omitted. The remaining LIDAR labels may be used as ground truth labels to generate ground truth data for a machine learning model (e.g., machine learning model 108).

[0100] Exemplary Autonomous Vehicle 13A is a diagram of an example autonomous vehicle 1300 according to some embodiments of the disclosure. Autonomous vehicle 1300 (alternatively referred to herein as “vehicle 1300”) may include, but is not limited to, a passenger vehicle, such as a car, a truck, a bus, a first responder vehicle, a shuttle, an electric or moped, a motorcycle, a fire engine, a police vehicle, an ambulance, a boat, a construction vehicle, a submarine, a drone, and / or another type of vehicle (e.g., unmanned and / or carrying one or more passengers). Autonomous vehicles are generally described in terms of automation levels 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) "Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles" (Standard No. J3016-201806 published June 15, 2018, Standard No. J3016-201609 published September 30, 2016, and previous and future versions of this standard). The mobile vehicle 1300 may be capable of functionality according to one or more of the autonomous driving levels Levels 3 through 5. For example, the mobile vehicle 1300 may be capable of conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5), depending on the embodiment.

[0101] The mobile vehicle 1300 may include components such as a chassis, body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of the mobile vehicle. The mobile vehicle 1300 may include a propulsion system 1350, such as an internal combustion engine, a hybrid power plant, a fully electric engine, and / or another propulsion system type. The propulsion system 1350 may be connected to a drive train of the mobile vehicle 1300, which may include a transmission, to enable propulsion of the mobile vehicle 1300. The propulsion system 1350 may be controlled in response to receiving a signal from a throttle / accelerator 1352.

[0102] A steering system 1354, which may include a steering wheel, may be used to steer the vehicle 1300 (e.g., along a desired course or route) when the propulsion system 1350 is operating (e.g., when the vehicle is moving). The steering system 1354 may receive signals from a steering actuator 1356. The steering wheel may be an option for fully automated (Level 5) functionality.

[0103] Brake sensor system 1346 may be used to operate vehicle brakes in response to receiving a signal from a brake actuator 1348 and / or a brake sensor.

[0104] A controller 1336, which may include one or more system on chip (SoC) 1304 (FIG. 13C) and / or a GPU, can provide signals (e.g., representations of commands) to one or more components and / or systems of the vehicle 1300. For example, the controller can send signals to operate vehicle brakes via one or more brake actuators 1348, operate a steering system 1354 via one or more steering actuators 1356, and operate a propulsion system 1350 via one or more throttle / accelerators 1352. The controller 1336 may include one or more on-board (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals and output operational commands (e.g., signals representing commands) to enable autonomous driving and / or to assist a driver in driving the vehicle 1300. The controllers 1336 may include a first controller 1336 for autonomous driving functions, a second controller 1336 for functional safety functions, a third controller 1336 for artificial intelligence functions (e.g., computer vision), a fourth controller 1336 for infotainment functions, a fifth controller 1336 for redundancy in emergency situations, and / or other controllers. In some instances, a single controller 1336 may handle two or more of the aforementioned functions, and two or more controllers 1336 may handle a single function, and / or any combination thereof.

[0105] Controller 1336 may provide signals to control one or more components and / or systems of vehicle 1300 in response to sensor data (e.g., sensor inputs) received from one or more sensors. The sensor data may be received from, for example, and without limitation, global navigation satellite system sensors 1358 (e.g., global positioning system sensors), RADAR sensors 1360, ultrasonic sensors 1362, LIDAR sensors 1364, inertial measurement unit (IMU) sensors 1366 (e.g., accelerometers, gyroscopes, magnetic compasses, magnetometers, etc.), microphones 1396, stereo cameras 1368, wide-view cameras 1370 (e.g., fisheye cameras), infrared cameras 1372, surround cameras 1374 (e.g., 360 degree cameras), long-range and / or medium-range cameras 1398, speed sensors 1344 (e.g., for measuring the speed of the moving vehicle 1300), vibration sensors 1342, steering sensors 1340, brake sensors (e.g., as part of a brake sensor system 1346), and / or other sensor types.

[0106] One or more of the controllers 1336 may receive input (e.g., represented by input data) from the instrument cluster 1332 of the vehicle 1300 and provide output (e.g., represented by output data, display data, etc.) via a human-machine interface (HMI) display 1334, an audible annunciator, a loudspeaker, and / or other components of the vehicle 1300. The output may include information such as vehicle velocity, speed, time, map data (e.g., HD map 1322 of FIG. 13C), position data (e.g., the position of the vehicle 1300 on a map, etc.), direction, positions of other vehicles (e.g., occupancy grid), information about objects and object situations as perceived by the controller 1336, etc. For example, the HMI display 1334 may display information regarding the presence of one or more objects (e.g., road signs, warning signs, changing traffic signals, etc.) and / or a driving maneuver that the moving vehicle has performed, is performing, or will perform (e.g., changing lanes now, taking exit 34B in 2 miles, etc.).

[0107] The mobile vehicle 1300 further includes a network interface 1324 capable of communicating over one or more networks using one or more wireless antennas 1326 and / or a modem. For example, the network interface 1324 may be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The wireless antennas 1326 may also enable communication between objects (e.g., mobile vehicles, mobile devices, etc.) in the environment using local area networks such as Bluetooth, Bluetooth LE, Z-Wave, ZigBee, etc., and / or low power wide-area networks (LPWANs) such as LoRaWAN, SigFox, etc.

[0108] 13B is an illustration of camera positions and fields of view of the example autonomous vehicle 1300 of FIG. 13A, according to some embodiments of the present disclosure. The cameras and their respective fields of view are one 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 positions on the vehicle 1300.

[0109] The camera type of the camera may include, but is not limited to, a digital camera that may be adapted for use with the components and / or systems of the mobile vehicle 1300. The camera may operate at automotive safety integrity level (ASIL) B and / or at another ASIL. The camera type may be capable of any image capture rate, such as 60 frames per second (fps), 1320 fps, 240 fps, etc., depending on the embodiment. The camera may be capable of using a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In some instances, the color filter array may include a red clear clear clear (RCCC) color filter array, a red clear clear blue (RCCB) color filter array, a red blue green clear (RBGC) color filter array, a Foveon X3 color filter array, a Bayer sensor (RGGB) color filter array, a monochrome sensor color filter array, and / or another type of color filter array. In some embodiments, clear pixel cameras, such as cameras with RCCC, RCCB, and / or RBGC color filter arrays, may be used in an effort to increase light sensitivity.

[0110] In some instances, one or more of the 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 assist, and intelligent headlamp control. One or more of the cameras (e.g., all of the cameras) may simultaneously record and provide image data (e.g., video).

[0111] One or more of the cameras may be mounted in a mounting part, such as a custom designed (3D printed) part, to filter out stray light and reflections from within the vehicle (e.g., reflections from the dashboard reflected in the windshield mirror) that may interfere with the camera's image data capture capabilities. With reference to a side mirror mounting part, the side mirror part may be custom 3D printed such that the camera mounting plate fits the shape of the side mirror. In some instances, the camera may be integrated into the side mirror. For side view cameras, the cameras may also be integrated into four posts at each corner of the cabin.

[0112] A camera (e.g., a forward-facing camera) with a field of view that includes a portion of the environment in front of the vehicle 1300 may be used for surround view to help identify the forward path and obstacles and, with the aid of one or more controllers 1336 and / or control SoCs, provide information essential to generating an occupancy grid and / or determining a preferred vehicle path. Forward-facing cameras may be used to perform many of the same ADAS functions as LIDAR, including emergency braking, pedestrian detection, and collision avoidance. Forward-facing cameras may also be used for ADAS functions and systems, including other functions such as lane departure warning ("LDW"), autonomous cruise control ("ACC"), and / or traffic sign recognition.

[0113] A variety of cameras may be used in a forward-facing configuration, including, for example, a monocular camera platform including a complementary metal oxide semiconductor (CMOS) color imager. Another example may be a wide-view camera 1370 that may be used to capture objects entering the view from the periphery (e.g., pedestrians, crossing traffic, or bicycles). Although only one wide-view camera is shown in FIG. 13B, there may be any number of wide-view cameras 1370 in the vehicle 1300. In addition, a long-range camera 1398 (e.g., a long-view stereo camera pair) may be used for depth-based object detection, especially for objects for which the neural network has not yet been trained. The long-range camera 1398 may also be used for object detection and classification, as well as basic object tracking.

[0114] One or more stereo cameras 1368 may also be included in the forward-facing configuration. The stereo camera 1368 may include an integrated control unit with an extensible processing unit that may provide programmable logic (FPGA) and a multi-core microprocessor with CAN or Ethernet interfaces integrated on a single chip. Such a unit may be used to generate a 3D map of the vehicle's environment, including distance estimates for all points in the image. An alternative stereo camera 1368 may include a compact stereo vision sensor that may include two camera lenses (one left and one right) and an image processing chip that can measure distances from the vehicle to objects of interest and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 1368 may be used in addition to or instead of those described herein.

[0115] A camera having a field of view that includes a portion of the environment to the side of the moving vehicle 1300 (e.g., a side-view camera) may be used for surround view, providing information used to create and update the occupancy grid and to generate side impact collision warnings. For example, surround cameras 1374 (e.g., four surround cameras 1374 as shown in FIG. 13B) may be positioned on the moving vehicle 1300. The surround cameras 1374 may include wide-view cameras 1370, fish-eye cameras, 360-degree cameras, and / or the like. For example, four fish-eye cameras may be positioned at the front, back, and sides of the moving vehicle. In an alternative arrangement, the moving vehicle may use three surround cameras 1374 (e.g., left, right, and rear) and may utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround view camera.

[0116] A camera having a field of view that includes the portion of the environment behind the moving vehicle 1300 (e.g., a rear-view camera) may be used for parking assistance, surround view, rear collision warning, and creating and updating an occupancy grid. As described herein, a wide variety of cameras may be used, including, but not limited to, cameras that are also suitable as forward-facing cameras (e.g., long-range and / or mid-range cameras 1398, stereo cameras 1368, infrared cameras 1372, etc.).

[0117] FIG. 13C is a block diagram of an example system architecture of the example autonomous vehicle 1300 of FIG. 13A, according to some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are merely illustrative. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, groupings of functions, etc.) may be used in addition to or instead of those shown, and some elements may be omitted altogether. Furthermore, many of the elements described herein are functional entities that may be implemented as separate or distributed components or in conjunction with other components, and in any suitable combination and location. Various functions described herein as being performed by an entity may be implemented by hardware, firmware, and / or software. For example, various functions may be implemented by a processor executing instructions stored in a memory.

[0118] Each of the components, features, and systems of the mobile vehicle 1300 of FIG. 13C are shown as connected via a bus 1302. The bus 1302 may include a controller area network (CAN) data interface (alternatively referred to as a "CAN bus"). The CAN may be a network within the mobile vehicle 1300 that is used to help control various features and functions of the mobile vehicle 1300, such as actuation of brakes, acceleration, braking, steering, windshield wipers, etc. The CAN bus may be configured to have dozens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus may be read to find steering angle, ground speed, engine revolutions per minute (RPM), button positions, and / or other vehicle status indicators. The CAN bus may be ASIL B compliant.

[0119] The bus 1302 is described herein as being a CAN bus, but this is not intended to be limiting. For example, FlexRay and / or Ethernet may be used in addition to or as an alternative to a CAN bus. Additionally, a single line is used to represent the bus 1302, but this is not intended to be limiting. There may be any number of buses 1302, which may include, for example, 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 instances, two or more buses 1302 may be used to perform different functions and / or may be used for redundancy. For example, a first bus 1302 may be used for collision avoidance functions and a second bus 1302 may be used for actuation control. In any instance, each bus 1302 may communicate with any of the components of the mobile vehicle 1300, and two or more buses 1302 may communicate with the same components. In some examples, each SoC 1304, each controller 1336, and / or each computer in the vehicle may have access to the same input data (e.g., input from sensors in the vehicle 1300) and may be connected to a common bus, such as a CAN bus.

[0120] The mobile vehicle 1300 may include one or more controllers 1336, such as those described herein with respect to FIGURE 13A. The controller 1336 may be used for a variety of functions. The controller 1336 may be coupled to any of a variety of other components and systems of the mobile vehicle 1300 and may be used for control of the mobile vehicle 1300, artificial intelligence of the mobile vehicle 1300, infotainment for the mobile vehicle 1300, and / or the like.

[0121] The mobile vehicle 1300 may include a system on a chip (SoC) 1304. The SoC 1304 may include a CPU 1306, a GPU 1308, a processor 1310, a cache 1312, an accelerator 1314, a data store 1316, and / or other components and features not shown. The SoC 1304 may be used to control the mobile vehicle 1300 in a variety of platforms and systems. For example, the SoC 1304 may be coupled in a system (e.g., the system of the mobile vehicle 1300) having an HD map 1322 that can obtain map refreshes and / or updates via a network interface 1324 from one or more servers (e.g., server 1378 of FIG. 13D).

[0122] CPU 1306 may include a CPU cluster or CPU complex (alternatively referred to as a "CCPLEX"). CPU 1306 may include multiple cores and / or L2 caches. For example, in some embodiments, CPU 1306 may include eight cores in a coherent multiprocessor configuration. In some embodiments, CPU 1306 may include four dual-core clusters, each with its own dedicated L2 cache (e.g., a 2M B L2 cache). CPU 1306 (e.g., CCPLEX) may be configured to support concurrent cluster operation allowing any combination of CPU 1306 clusters to be active at any given time.

[0123] The CPU 1306 may implement power management capabilities including one or more of the following features: individual hardware blocks may be automatically clock gated when idle to conserve dynamic power, each core clock may be gated when the core is not actively executing instructions by execution of 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. The CPU 1306 may further implement an enhanced algorithm for managing power states where allowed power states and expected wake-up times are specified and the hardware / microcode determines the best power state for entering the cores, clusters, and CCPLEXes. The processing cores may support simplified power state entry sequences in software where work is offloaded to the microcode.

[0124] The GPU 1308 may include an integrated GPU (alternatively referred to herein as an "iGPU"). The GPU 1308 may be programmable and efficient for parallel workloads. In some instances, the GPU 1308 may use an enhanced tensor instruction set. The GPU 1308 may include one or more streaming microprocessors, where each streaming microprocessor may include an L1 cache (e.g., an L1 cache having at least 96 KB storage capacity) and two or more of the streaming microprocessors may share a cache (e.g., an L2 cache having 512 KB storage capacity). In some embodiments, the GPU 1308 may include at least eight streaming microprocessors. The GPU 1308 may use a computational application programming interface (API). In addition, the GPU 1308 may use one or more parallel computing platforms and / or programming models (e.g., NVIDIA's CUDA).

[0125] The GPU 1308 may be power optimized for best performance in automotive and embedded use cases. For example, the GPU 1308 may be fabricated on FinFETs (Fin field-effect transistors). However, this is not intended to be limiting and the GPU 1308 may be fabricated using other semiconductor fabrication processes. Each streaming microprocessor may incorporate several mixed-precision processing cores partitioned into multiple blocks. For example, and without limitation, 64 PF32 cores and 32 PF64 cores may be partitioned into four processing blocks. In such an example, each processing block may be assigned 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed-precision NVIDIA tensor cores for deep learning matrix operations, an L0 instruction cache, a warp scheduler, a dispatch unit, and / or a 64KB register file. In addition, the streaming microprocessor may include independent parallel integer and floating-point data paths to provide efficient execution of workloads having a mix of computational and addressing operations. Streaming microprocessors may include independent thread scheduling capabilities to allow finer grain synchronization and coordination among concurrent threads. Streaming microprocessors may include a combined L1 data cache and shared memory unit to improve performance while simplifying programming.

[0126] The GPU 1308 may include high bandwidth memory (HBM) and / or a 16GB HBM2 memory subsystem to provide about 900GB / sec peak memory bandwidth in some instances. In some instances, 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 place of the HBM memory.

[0127] The GPU 1308 may include unified memory technology including access counters to enable more accurate movement of memory pages to the processors that access them most frequently, thereby improving the efficiency of storage ranges shared between processors. In some instances, address translation service (ATS) support may be used to enable the GPU 1308 to directly access the CPU 1306 page tables. In such instances, when the GPU 1308 memory management unit (MMU) experiences a miss, an address translation request may be sent to the CPU 1306. In response, the CPU 1306 may consult its page table for a virtual-to-real mapping of the address and send the translation back to the GPU 1308. As such, the unified memory technology may enable a single unified virtual address space for both the CPU 1306 and GPU 1308 memories, thereby simplifying GPU 1308 programming and porting of applications to the GPU 1308.

[0128] In addition, the GPU 1308 may include access counters that can record the frequency of accesses of the GPU 1308 to the memory of other processors. The access counters can help ensure that memory pages are moved to the physical memory of the processor that is accessing the page most frequently.

[0129] The SoC 1304 may include any number of caches 1312, including those described herein. For example, the cache 1312 may include an L3 cache available to both the CPU 1306 and the GPU 1308 (e.g., connected to both the CPU 1306 and the GPU 1308). The cache 1312 may include a write-back cache that can record line state, such as by using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). The L3 cache may include 4 MB or more, depending on the embodiment, although smaller cache sizes may be used.

[0130] The SoC 1304 may include one or more accelerators 1314 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, the SoC 1304 may include a hardware acceleration cluster, which may include optimized hardware accelerators and / or large on-chip memory. The large on-chip memory (e.g., 4MB of SRAM) may enable the hardware acceleration cluster to accelerate neural networks and other operations. The hardware acceleration cluster may be used to complement the GPU 1308 and to offload some of the tasks of the GPU 1308 (e.g., to free up more cycles for the GPU 1308 to perform other tasks). As an example, the accelerator 1314 may be used for target workloads that are stable enough to be suitable for acceleration (e.g., perception, convolutional neural networks (CNNs), etc.). As used herein, the term "CNN" may include all types of CNNs, including region-based or regional convolutional neural networks (RCNNs) and Faster RCNNs (e.g., as used for object detection).

[0131] The accelerator 1314 (e.g., a hardware acceleration cluster) may include a deep learning accelerator (DLA). The DLA may include one or more tensor processing units (TPUs) that can be configured to provide an additional 10 trillion operations per second for deep learning applications and inference. The TPU may be an accelerator configured and optimized to perform image processing functions (e.g., CNN, RCNN, etc.). The DLA may further be optimized for a specific set of neural network types and floating point operations, as well as inference. The design of the DLA may provide more performance per millimeter than a general-purpose GPU, greatly exceeding the performance of a CPU. The TPU may perform several functions, including, for example, single instance convolution functions, supporting INT8, INT16, and FP16 data types for both features and weights, as well as post-processor functions.

[0132] The DLA can quickly and efficiently run neural networks, in particular CNNs, on processed or unprocessed data for any of a variety of functions, including but not limited to: CNNs for object identification and detection using data from camera sensors, CNNs for distance estimation using data from camera sensors, CNNs for emergency vehicle detection and identification using data from microphones, CNNs for face recognition and moving vehicle owner identification using data from camera sensors, and / or CNNs for security and / or safety related events.

[0133] The DLA can perform any function of the GPU 1308, and by using an inference accelerator, for example, a designer can target either the DLA or the GPU 1308 for any function. For example, a designer can focus on processing CNNs and floating point operations on the DLA and offload other functions to the GPU 1308 and / or other accelerators 1314.

[0134] The accelerator 1314 (e.g., hardware acceleration cluster) may include a programmable vision accelerator (PVA), which may alternatively be referred to herein as a computer vision accelerator. The PVA 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. The PVA may provide a balance between performance and flexibility. For example, each PVA may include, but is not limited to, any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors, for example.

[0135] The RISC cores may interact with an image sensor (e.g., an image sensor of any of the cameras described herein), an image signal processor, and / or the like. Each RISC core may include any amount of memory. The RISC cores may use any of a number of protocols, depending on the embodiment. In some instances, the RISC cores may run a real-time operating system (RTOS). The RISC cores may be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC cores may include an instruction cache and / or tightly coupled RAM.

[0136] The DMA may allow components of the PVA to access system memory independent of the CPU 1306. The DMA may support any number of features used to provide optimizations to the PVA, including but not limited to supporting multi-dimensional addressing and / or circular addressing. In some instances, 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.

[0137] The vector processor may be a programmable processor that may be designed to efficiently and flexibly execute computer vision algorithm programming and provide signal processing capabilities. In some instances, the PVA may include a PVA core and two vector processing subsystem partitions. The PVA core may include a processor subsystem, a DMA engine (e.g., two DMA engines), and / or other peripherals. The vector processing subsystem may act as the primary 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, for example, a single instruction, multiple data (SIMD), very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW may increase throughput and speed.

[0138] Each vector processor may include an instruction cache and may be coupled to dedicated memory. As a result, in some instances, each vector processor may be configured to execute independently of other vector processors. In other instances, the vector processors included in a particular PVA may be configured to employ data parallelism. For example, in some embodiments, multiple vector processors included in a single PVA may execute the same computer vision algorithm, but on different regions of an image. In other instances, the vector processors included in a particular PVA may execute different computer vision algorithms simultaneously on the same image, or even execute different algorithms on sequential images or portions of an image. In particular, any number of PVAs may be included in a hardware accelerated cluster, and any number of vector processors may be included in each PVA. In addition, the PVAs may include additional error correcting code (ECC) memory to increase overall system security.

[0139] The accelerator 1314 (e.g., a hardware acceleration cluster) may include a computer vision network-on-chip and SRAM to provide high bandwidth, low latency SRAM for the accelerator 1314. In some instances, the on-chip memory may include, for example and without limitation, at least 4 MB of SRAM consisting of eight field configurable memory blocks that may be accessible by both the PVA and DLA. Each pair of memory blocks may include an advanced peripheral bus (APB) interface, configuration circuitry, controllers, and multiplexers. Any type of memory may be used. The PVA and DLA may access the memory through a backbone that provides the PVA and DLA with high speed access to the memory. The backbone may include a computer vision network-on-chip that interconnects the PVA and DLA to the memory (e.g., using the APB).

[0140] The computer vision network-on-chip may include an interface that determines, prior to transmission of any control signals / addresses / data, that both the PVA and DLA provide ready and valid signals. Such an interface may provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst type communication for continuous data transfer. This type of interface may follow ISO26262 or IEC61508 standards, although other standards and protocols may be used.

[0141] In some instances, the SoC 1304 may include a real-time ray tracing hardware accelerator, such as described in U.S. Patent Application No. 16 / 101,232, filed Aug. 10, 2018. The real-time ray tracing hardware accelerator may be used to quickly and efficiently determine the location and scale of objects (e.g., within a world model) to generate real-time visualization simulations for RADAR signal interpretation, for acoustic propagation synthesis and / or analysis, for simulation of SONAR systems, for general wave propagation simulation, for comparison against LIDAR data for localization and / or other functions, and / or for other uses.

[0142] The accelerator 1314 (e.g., a hardware accelerator cluster) has a variety of applications for autonomous driving. The PVA may be a programmable vision accelerator that can be used for critical processing stages in ADAS and autonomous vehicles. The capabilities of the PVA are well suited to the domain of algorithms that require predictable processing at low power and low latency. In other words, the PVA works well with semi-dense or dense regular computations on small data sets that require predictable execution times with low latency and low power. Hence, in the context of a platform for autonomous vehicles, the PVA is designed to run classical computer vision algorithms, since the PVA is efficient in object detection and operating on integer computations.

[0143] For example, according to one embodiment of the present technology, the PVA is used to perform computer stereo vision. A semi-global matching based algorithm may be used in some instances, but this is not intended to be limiting. Many applications for Level 3-5 autonomous driving require motion estimation / stereo matching on the fly (e.g., structure from motion, pedestrian recognition, lane detection, etc.). The PVA can perform computer stereo vision functions with input from two monocular cameras.

[0144] In some instances, PVA may be used to perform dense optical flow by processing raw RADAR data (e.g., using a 4D Fast Fourier Transform) to provide processed RADAR. In other instances, 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.

[0145] DLA can be used to implement any type of network to enhance control and driving safety, including, for example, a neural network that outputs a measure of the confidence of each object detection. Such confidence value can be interpreted as a probability or as providing a relative "weight" of each detection compared to other detections. This confidence value allows the system to make further decisions regarding which detections should be considered true positive detections rather than false positive detections. For example, the system can set a confidence threshold and consider only detections above the threshold as true positive detections. In an automatic emergency braking (AEB) system, a false positive detection would cause a moving vehicle to automatically perform emergency braking, which is clearly undesirable. Therefore, only the most confident detections should be considered to trigger AEB. DLA can implement a neural network that regresses the confidence value. The neural network may receive as its inputs at least some subset of parameters, such as bounding box dimensions, ground plane estimates obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 1366 outputs that correlate with the vehicle 1300 orientation, distance, and 3D position estimates of the object obtained from the neural network and / or other sensors (e.g., LIDAR sensor 1364 or RADAR sensor 1360), and others.

[0146] The SoC 1304 may include a data store 1316 (e.g., a memory). The data store 1316 may be an on-chip memory of the SoC 1304 and may store the neural network to be executed by the GPU and / or DLA. In some instances, the data store 1316 may have a capacity large enough to store multiple instances of the neural network for redundancy and safety. The data store 1312 may comprise an L2 or L3 cache 1312. References to the data store 1316 may include references to memory associated with the PVA, DLA, and / or other accelerators 1314, as described herein.

[0147] The SoC 1304 may include one or more processors 1310 (e.g., embedded processors). The processors 1310 may include a boot and power management processor, which may be a dedicated processor and subsystem for handling boot power and management capabilities and associated security enforcement. The boot and power management processor may be part of the SoC 1304 boot sequence and may provide run-time power management services. The boot power and management processor may provide clock and voltage programming, assist with system low power state transitions, management of the SoC 1304 thermal and temperature sensors, and / or management of the SoC 1304 power states. Each temperature sensor may be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 1304 may use the ring oscillator to detect the temperature of the CPU 1306, the GPU 1308, and / or the accelerator 1314. If the temperature is determined to exceed a threshold, the boot and power management processor may enter a temperature fault routine, place the SoC 1304 in a lower power state and / or place the vehicle 1300 in a Chauffeur safe shutdown mode (e.g., bring the vehicle 1300 to a safe shutdown).

[0148] The processor 1310 may further include a set of embedded processors that may perform the functions of an audio processing engine. The audio processing engine may be an audio subsystem that allows full hardware support of multi-channel audio through multiple interfaces and a wide and flexible range of audio I / O interfaces. In some instances, the audio processing engine is a dedicated processor core that includes a digital signal processor with dedicated RAM.

[0149] The processor 1310 may further include an always-on processor engine that can provide the necessary hardware features to support low power sensor management and wake use cases. The always-on processor engine may include a processor core, tightly coupled RAM, support peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0150] The processor 1310 may further include a safety cluster engine that includes a processor subsystem dedicated to handling safety management of 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 a safety mode, the two or more cores may operate in lockstep mode and function as a single core with comparison logic to detect any differences between their operations.

[0151] The processor 1310 may further include a real-time camera engine, which may include a dedicated processor subsystem for handling real-time camera management.

[0152] The processor 1310 may further 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 pipeline.

[0153] The processor 1310 may include a video image compositor, which may be a processing block (e.g., implemented in a microprocessor) that implements video post-processing functions required by the video playback application to produce a final image for the player window. The video image compositor may perform lens distortion correction on the wide-view camera 1370, on the surround camera 1374, and / or on the in-cabin surveillance camera sensor. The in-cabin surveillance camera sensor is preferably monitored by a neural network running on another instance of the advanced SoC, configured to identify in-cabin events and respond appropriately. The in-cabin system may perform lip reading to activate cellular service and make phone calls, dictate emails, change the destination of the vehicle, activate or change the vehicle's infotainment system and settings, or provide voice-activated web surfing. Certain features are only available to the driver when operating in autonomous mode and are disabled otherwise.

[0154] The video image synthesizer may include enhanced temporal noise reduction for both spatial and temporal noise reduction. For example, when motion occurs in the video, the noise reduction reduces the weight of information provided by adjacent frames and appropriately weights the spatial information. When an image or part of an image does not contain motion, the temporal noise reduction performed by the video image synthesizer may use information from previous images to reduce noise in the current image.

[0155] The video image composer may also be configured to perform stereo rectification on the input stereo lens frames. The video image composer may further be used for user interface compositing when the operating system desktop is in use and the GPU 1308 is not required to continuously render new surfaces. Even when the GPU 1308 is powered on and actively performing 3D rendering, the video image composer may be used to offload the GPU 1308 for improved performance and responsiveness.

[0156] The SoC1304 may further include a mobile industry processor interface (MIPI) camera serial interface, a high-speed interface for receiving video and input from a camera, and / or a video input block that may be used for camera and related pixel input functions. The SoC1304 may further include an input / output controller that may be controlled by software and that may be used to receive I / O signals that are not committed to a specific role.

[0157] The SoC 1304 may further include a wide range of peripheral interfaces to enable communication with peripherals, audio codecs, power management, and / or other devices. The SoC 1304 may be used to process data from cameras (e.g., connected via gigabit multimedia serial links and Ethernet), sensors (e.g., LIDAR sensor 1364, RADAR sensor 1360, etc., which may be connected via Ethernet), data from the bus 1302 (e.g., vehicle 1300 speed, steering wheel position, etc.), and GNSS sensor 1358 (e.g., connected via Ethernet or CAN bus). The SoC 1304 may further include a dedicated high-performance mass storage controller, which may include its own DMA engine and may be used to offload the CPU 1306 from routine data management tasks.

[0158] The SoC1304 may be an end-to-end platform with a flexible architecture that spans levels 3-5 of automation, thereby providing a comprehensive functional safety architecture that leverages and efficiently uses computer vision and ADAS techniques for diversity and redundancy, and provides a platform for a flexible, reliable driving software stack along with deep learning tools. The SoC1304 may be faster, more reliable, and more energy- and space-efficient than traditional systems. For example, when the accelerator 1314 is combined with the CPU 1306, the GPU 1308, and the data store 1316, the SoC1304 may provide a fast and efficient platform for levels 3-5 of autonomous vehicles.

[0159] Thus, the present technology provides capabilities and functionality that cannot be achieved by conventional systems. For example, computer vision algorithms may be executed on a CPU, which may be configured using a high-level programming language, such as the C programming language, to execute a wide variety of processing algorithms across a wide variety of visual data. However, CPUs often cannot meet the performance requirements of many computer vision applications, such as those related to execution time and power consumption. In particular, many CPUs cannot execute complex object detection algorithms in real time, which is a requirement for in-vehicle ADAS applications and for practical Level 3-5 autonomous vehicles.

[0160] In contrast 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 run simultaneously and / or sequentially and the results to be combined to enable Level 3-5 autonomous driving capabilities. For example, a CNN running on the DLA or dGPU (e.g., GPU1320) can include text and word recognition that allows the supercomputer to read and understand traffic signs, including signs for which the neural network was not specifically trained. The DLA can further include neural networks that can identify, interpret, and provide a semantic understanding of the signs and pass the semantic understanding to a route planning module running on the CPU complex.

[0161] As another example, multiple neural networks may be run simultaneously, as required for level 3, 4, or 5 operation. For example, a warning sign consisting of "Caution: Flashing lights indicate icy conditions" along with a lightning flash may be interpreted by several neural networks, independently or collectively. The sign itself may be identified as a traffic sign by a first deployed neural network (e.g., a trained neural network), and the text "Flashing lights indicate icy conditions" may be interpreted by a second deployed neural network that informs the vehicle's route planning software (preferably running on a CPU complex) that icy conditions exist when a flashing light is detected. The flashing light may be identified by running a third deployed neural network through multiple frames, informing the vehicle's route planning software of the presence (or absence) of the flashing light. All three neural networks may be run simultaneously, such as within the DLA and / or on the GPU 1308.

[0162] In some instances, a CNN for facial recognition and vehicle owner identification can use data from the camera sensor to identify the presence of a legitimate driver and / or owner of the vehicle 1300. The always-on sensor processing engine can be used to unlock the vehicle and turn on the lights when the owner approaches the driver's door, and in security mode, to stop operation of the vehicle when the owner leaves the vehicle. In this manner, the SoC 1304 provides security against theft and / or vehicle takeover.

[0163] In another example, a CNN for emergency vehicle detection and identification can detect and identify emergency vehicle sirens using data from the microphone 1396. In contrast to conventional systems that use general classifiers to detect sirens and manually extract features, the SoC 1304 uses CNNs for classification of environmental and urban sounds, as well as classification of visual data. In a preferred embodiment, the CNN running on the DLA is trained to identify the relative terminal velocity 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 mobile vehicle is operating, as identified by the GNSS sensor 1358. Thus, for example, when operating in Europe, the CNN will attempt to detect European sirens, and when in the United States, the CNN will attempt to identify only North American sirens. After an emergency vehicle is detected, the control program may be used to execute emergency vehicle safety routines, such as slowing down the mobile vehicle, stopping the mobile vehicle at the side of the road, parking the mobile vehicle, and / or idling the mobile vehicle, with the assistance of ultrasonic sensor 1362, until the emergency vehicle has passed.

[0164] The vehicle may include a CPU 1318 (e.g., a separate CPU, or a dCPU) that may be coupled to the SoC 1304 via a high-speed interconnect (e.g., PCIe). The CPU 1318 may include, for example, an X86 processor. The CPU 1318 may be used to perform any of a variety of functions including, for example, reconciling potentially inconsistent results between the ADAS sensors and the SoC 1304 and / or monitoring the status and health of the controller 1336 and / or the infotainment SoC 1330.

[0165] The vehicle 1300 may include a GPU 1320 (e.g., a discrete GPU, or a dGPU) that may be coupled to the SoC 1304 via a high-speed interconnect (e.g., NVIDIA's NVLINK). The GPU 1320 may provide additional artificial intelligence functionality, such as by running redundant and / or different neural networks, and may be used to train and / or update neural networks based on input (e.g., sensor data) from sensors of the vehicle 1300.

[0166] The mobile vehicle 1300 may further include a network interface 1324 that may include one or more wireless antennas 1326 (e.g., one or more wireless antennas for different communication protocols, such as a cellular antenna, a Bluetooth antenna, etc.). The network interface 1324 may be used to enable wireless connections with the cloud over the Internet (e.g., with the server 1378 and / or other network devices), with other mobile vehicles, and / or with computing devices (e.g., passenger client devices). To communicate with other mobile vehicles, a direct link may be established between the two mobile vehicles and / or an indirect link may be established (e.g., through a network and via the Internet). A direct link may be provided using a mobile vehicle-to-mobile communication link. The mobile vehicle-to-mobile communication link may provide the mobile vehicle 1300 information about mobile vehicles in its vicinity (e.g., mobile vehicles in front of, beside, and / or behind the mobile vehicle 1300). This feature may be part of a collaborative adaptive cruise control feature of the vehicle 1300 .

[0167] The network interface 1324 may include a SoC that provides modulation and demodulation functions and enables the controller 1336 to communicate over a wireless network. The network interface 1324 may include a radio frequency front end for upconversion from baseband to radio frequency and downconversion from radio frequency to baseband. The frequency conversion may be performed through well known processes and / or may be performed using a superheterodyne process. In some instances, the radio frequency front end functions may be provided by a separate chip. The network interface may include wireless functions for communicating via LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0168] Mobile vehicle 1300 may further include a data store 1328, which may include off-chip (e.g., off-SoC 1304) storage. Data store 1328 may include one or more memory elements, including RAM, SRAM, DRAM, VRAM, flash, hard disk, and / or other components and / or devices capable of storing at least one bit of data.

[0169] The vehicle 1300 may further include a GNSS sensor 1358 (e.g., a GPS and / or an assisted GPS sensor) to assist with mapping, perception, occupancy grid generation, and / or route planning functions. Any number of GNSS sensors 1358 may be used, including, but not limited to, a GPS using a USB connector with an Ethernet to serial (RS-232) bridge.

[0170] The mobile vehicle 1300 may further include a RADAR sensor 1360. The RADAR sensor 1360 may be used by the mobile vehicle 1300 for long-range mobile vehicle detection, even in darkness and / or severe weather conditions. The RADAR functional safety level may be ASIL B. In some instances, the RADAR sensor 1360 may use the CAN and / or bus 1302 for control and to access object tracking data (e.g., to transmit data generated by the RADAR sensor 1360), with access to Ethernet to access raw data. A wide variety of RADAR sensor types may be used. For example, and without limitation, the RADAR sensor 1360 may be suitable for front, rear, and side RADAR use. In some instances, a pulsed Doppler RADAR sensor is used.

[0171] The RADAR sensor 1360 may include different configurations, such as long range with a narrow field of view, short range with a wide field of view, and short range side coverage. In some instances, the long range RADAR may be used for adaptive cruise control functions. Long range RADAR systems may provide a wide field of view achieved by two or more independent scans, such as within a 250m range. The RADAR sensor 1360 may help distinguish between static and moving objects and may be used by ADAS systems for emergency brake assist and forward collision warning. Long range RADAR sensors may include monostatic multimodal RADAR with multiple (e.g., six or more) fixed RADAR antennas and high speed CAN and FlexRay interfaces. In one instance with six antennas, the center four antennas may create a focused beam pattern designed to record the surroundings of the moving vehicle 1300 at high speeds with minimal interference from traffic in adjacent lanes. The other two antennas may widen the field of view, allowing for rapid detection of moving vehicles entering or leaving the lane of the moving vehicle 1300.

[0172] As an example, a medium-range RADAR system may include a range of up to 1360m (front) or 80m (rear) and a field of view of up to 42 degrees (front) or 1350 degrees (rear). A short-range RADAR system may include, but is not limited to, a RADAR sensor designed to be mounted on either end of a rear bumper. When mounted on either end of a rear bumper, such a RADAR sensor system can create two beams that constantly monitor the blind spots behind and adjacent to a moving vehicle.

[0173] Short-range RADAR systems may be used in ADAS systems for blind spot detection and / or lane change assist.

[0174] The mobile vehicle 1300 may further include ultrasonic sensors 1362. The ultrasonic sensors 1362, which may be located at the front, rear, and / or sides of the mobile vehicle 1300, may be used for parking assist and / or for creating and updating an occupancy grid. A wide variety of ultrasonic sensors 1362 may be used, with different ultrasonic sensors 1362 being used for different ranges of detection (e.g., 2.5m, 4m). The ultrasonic sensors 1362 may operate at a functional safety level of ASIL B.

[0175] The mobile vehicle 1300 may include a LIDAR sensor 1364. The LIDAR sensor 1364 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensor 1364 may be functional safety level ASIL B. In some instances, the mobile vehicle 1300 may include multiple (e.g., 2, 4, 6, etc.) LIDAR sensors 1364 that may use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0176] In some instances, the LIDAR sensor 1364 may have the ability to provide a list of objects and their distances in a 360-degree field of view. A commercially available LIDAR sensor 1364 may have an advertised range of about 1300 m, for example, with an accuracy of 2 cm to 3 cm, and with support for a 1300 Mbps Ethernet connection. In some instances, one or more non-protruding LIDAR sensors 1364 may be used. In such instances, the LIDAR sensor 1364 may be implemented as a small device that may be integrated into the front, rear, sides, and / or corners of the vehicle 1300. In such instances, the LIDAR sensor 1364 may have a range of 200 m even for low-reflecting objects and provide up to a 1320-degree horizontal and 35-degree vertical field of view. A front-mounted LIDAR sensor 1364 may be configured for a horizontal field of view between 45 and 135 degrees.

[0177] In some instances, LIDAR technology such as 3D flash LIDAR may also be used. 3D flash LIDAR uses a laser flash as a transmitter to illuminate the surroundings of the vehicle up to about 200 m. The flash LIDAR unit includes a receptor that records the laser pulse transit time and the reflected light on each pixel, which in turn corresponds to the range from the vehicle to the object. Flash LIDAR may allow a highly accurate and distortion-free image of the surroundings to be generated with every laser flash. In some instances, four flash LIDAR sensors may be deployed, one on each side of the vehicle 1300. Available 3D flash LIDAR systems include solid-state 3D steering array LIDAR cameras (e.g., non-scanning LIDAR devices) that have no moving parts other than the blower. The flash LIDAR device may use a 5 nanosecond Class I (eye-safe) laser pulse per frame and may capture the reflected laser light in the form of a 3D range point cloud and coregistered intensity data. By using flash LIDAR, and because flash LIDAR is a solid-state device with no moving parts, the LIDAR sensor 1364 may be less susceptible to motion blur, vibration, and / or shock.

[0178] The vehicle may further include an IMU sensor 1366. In some instances, the IMU sensor 1366 may be positioned at the center of the rear axle of the vehicle 1300. The IMU sensor 1366 may include, for example, but not limited to, an accelerometer, a magnetometer, a gyroscope, a magnetic compass, and / or other sensor types. In some instances, such as in a 6-axis application, the IMU sensor 1366 may include an accelerometer and a gyroscope, while in a 9-axis application, the IMU sensor 1366 may include an accelerometer, a gyroscope, and a magnetometer.

[0179] In some embodiments, the IMU sensor 1366 may be implemented as a miniature, high-performance GPS-Aided Inertial Navigation System (GPS / INS) that combines micro-electro-mechanical system (MEMS) inertial sensors, highly sensitive GPS receivers, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. As such, in some instances, the IMU sensor 1366 may enable the vehicle 1300 to estimate heading without requiring input from a magnetic sensor by directly observing and correlating changes in velocity from the GPS to the IMU sensor 1366. In some instances, the IMU sensor 1366 and the GNSS sensor 1358 may be combined in a single integrated unit.

[0180] The vehicle may include microphones 1396 placed in and / or around the vehicle 1300. The microphones 1396 may be used for emergency vehicle detection and identification, among other things.

[0181] The vehicle may further include any number of camera types, including stereo cameras 1368, wide view cameras 1370, infrared cameras 1372, surround cameras 1374, long and / or mid-range cameras 1398, and / or other camera types. The cameras may be used to capture image data around the entire exterior of the vehicle 1300. The type of cameras used depends on the embodiment and requirements of the vehicle 1300, and any combination of camera types may be used to achieve the required coverage around the vehicle 1300. Additionally, the number of cameras may vary depending on the embodiment. For example, the vehicle may include six cameras, seven cameras, ten cameras, twelve cameras, and / or another number of cameras. The cameras may support, by way of example only, Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet, but are not limited thereto. Each camera is described in further detail herein in conjunction with Figures 13A and 13B.

[0182] The mobile vehicle 1300 may further include a vibration sensor 1342. The vibration sensor 1342 may measure vibrations of components of the mobile vehicle, such as an axle. For example, a change in vibration may indicate a change in the surface of the road. In another example, when two or more vibration sensors 1342 are used, the difference in vibration may be used to determine friction or slippage of the road surface (e.g., when the difference in vibration is between a powered axle and a free-running axle).

[0183] The mobile vehicle 1300 may include an ADAS system 1338. In some instances, the ADAS system 1338 may include a SoC. The ADAS system 1338 may include autonomous / adaptive / automatic cruise control (ACC), cooperative adaptive cruise control (CACC), forward crash warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keep assist (LKA), blind spot warning (BSW), rear cross-traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functions.

[0184] The ACC system may use a RADAR sensor 1360, a LIDAR sensor 1364, and / or a camera. The ACC system may include longitudinal ACC and / or lateral ACC. The longitudinal ACC monitors and controls the distance to the vehicle directly ahead of the vehicle 1300 and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle ahead. The lateral ACC performs distance keeping and advises the vehicle 1300 to change lanes when necessary. The lateral ACC is related to other ADAS applications such as LCA and CWS.

[0185] CACC uses information from other moving vehicles, which may be received from other moving vehicles via a wireless link via the network interface 1324 and / or the wireless antenna 1326, or indirectly via a network connection (e.g., via the Internet). A direct link may be provided by a vehicle-to-vehicle (V2V) communication link, while an indirect link may be an infrastructure-to-vehicle (I2V) communication link. In general, a V2V communication concept provides information about the immediately preceding moving vehicle (e.g., the moving vehicle directly ahead of the moving vehicle 1300, in the same lane as the moving vehicle 1300), while an I2V communication concept provides information about the traffic further ahead. A CACC system may include either or both I2V and V2V information sources. Given the information of moving vehicles ahead of the moving vehicle 1300, CACC may be more reliable and may potentially make traffic flow smoother and reduce congestion on the roads.

[0186] The FCW system is designed to warn the driver of hazards so that the driver can take corrective action. The FCW system uses a forward-facing camera and / or RADAR sensor 1360 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, electrically coupled to driver feedback such as a display, speaker, and / or vibrating components. The FCW system can provide warnings in the form of audio, visual alerts, vibrations, and / or quick brake pulses.

[0187] An AEB system can detect an imminent forward collision with another moving vehicle or other object and automatically apply the brakes if the driver does not take corrective action within specified time or distance parameters. The AEB system can use a forward-facing camera and / or RADAR sensor 1360 coupled to a dedicated processor, DSP, FPGA, and / or ASIC. When the AEB system detects a hazard, the AEB system typically first alerts the driver to take corrective action to avoid the collision, and if the driver does not take corrective action, the AEB system can automatically apply the brakes as part of an effort to prevent or at least mitigate the effects of the predicted collision. The AEB system can include techniques such as dynamic brake support and / or collision imminent braking.

[0188] The LDW system provides visual, audible, and / or tactile warnings, such as vibrations of the steering wheel or seat, to alert the driver when the mobile vehicle 1300 crosses a lane marking. The LDW system is not activated when the driver indicates an intentional lane departure by activating a turn signal. The LDW system may use a forward-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC electrically coupled to driver feedback, such as a display, speaker, and / or vibrating components.

[0189] The LKA system is a modification of the LDW system, which provides steering input or braking to correct the vehicle 1300 if it begins to wander out of its lane.

[0190] The BSW system detects and warns the driver of a moving vehicle in the vehicle's blind spot. The BSW system can provide visual, audible, and / or tactile warnings to indicate that merging or changing lanes is unsafe. The system can provide additional warnings when the driver uses turn signals. The BSW system can use rear-facing cameras and / or RADAR sensors 1360 coupled to dedicated processors, DSPs, FPGAs, and / or ASICs electrically coupled to driver feedback, e.g., displays, speakers, and / or vibration components.

[0191] The RCTW system may provide visual, audible, and / or tactile notifications when an object is detected outside the range of the rear camera when the vehicle 1300 is backing up. Some RCTW systems include AEB to ensure vehicle brakes are applied to avoid a collision. The RCTW system may use one or more rear-facing RADAR sensors 1360 coupled to a dedicated processor, DSP, FPGA, and / or ASIC electrically coupled to driver feedback, e.g., a display, speaker, and / or vibration components.

[0192] Because conventional ADAS systems alert the driver and allow the driver to determine whether a safety condition truly exists and act accordingly, conventional ADAS systems can be prone to producing false positive results that are usually not catastrophic but can be annoying and distracting to the driver. However, in an autonomous vehicle 1300, the vehicle 1300 itself must decide whether to listen to the results from the primary computer or the secondary computer (e.g., the first controller 1336 or the second controller 1336) when the results are conflicting. For example, in some embodiments, the ADAS system 1338 may be a backup and / or secondary computer to provide perception information to a backup computer rationality module. The backup computer rationality monitor can run redundant and diverse software on hardware components to detect failures in perception and dynamic driving tasks. Output from the ADAS system 1338 may be provided to a supervisory MCU. When the outputs from the primary and secondary computers conflict, the supervisory MCU must decide how to reconcile the conflict to ensure safe operation.

[0193] In some instances, the primary computer may be configured to provide a reliability score to the overseer MCU indicating the reliability of the primary computer in a selected outcome. If the reliability score exceeds a threshold, the overseer MCU may follow the instructions of the primary computer regardless of whether the secondary computers provide conflicting or inconsistent results. If the reliability score does not meet the threshold, and if the primary and secondary computers show different (e.g., conflicting) results, the overseer MCU may arbitrate between the computers to determine the appropriate outcome.

[0194] The supervisory MCU may be configured to execute a neural network that is trained and configured to determine conditions under which the secondary computer provides a false alarm based on the outputs from the primary and secondary computers. Thus, the neural network in the supervisory MCU may learn when the output of the secondary computer may be trusted and when it may not be trusted. For example, when the secondary computer is a RADAR-based FCW system, the neural network in the supervisory MCU may learn when the FCW identifies a metal object that is not actually a hazard, such as a sewer grate or manhole cover, that triggers an alarm. Similarly, when the secondary computer is a camera-based LDW system, the neural network in the supervisory MCU may learn to ignore the LDW when a bicyclist or pedestrian is present and lane departure is, in fact, the safest maneuver. In embodiments that include a neural network executing on the supervisory MCU, the supervisory MCU may include at least one of a DLA or a GPU suitable for executing the neural network with associated memory. In a preferred embodiment, the director MCU may comprise and / or be included as a component of the SoC1304.

[0195] In other instances, the ADAS system 1338 may include a secondary computer that performs ADAS functions using classical rules of computer vision. As such, the secondary computer may use classical computer vision rules (if-then) and the presence of a neural network in the supervisory MCU may improve reliability, safety and performance. For example, the diverse implementations and intentional non-identity may make the overall system more fault-tolerant, especially to failures caused by software (or software-hardware interface) functions. For example, if there is a software bug or error in the software running on the primary computer and non-identical software code running on the secondary computer provides the same overall result, the supervisory MCU may have greater confidence that the overall result is correct and that a bug in the software or hardware on the primary computer has not caused a critical error.

[0196] In some instances, the output of the ADAS system 1338 may be fed to a perception block of the primary computer and / or a dynamic driving task block of the primary computer. For example, if the ADAS system 1338 indicates a forward collision warning due to an object directly ahead, the perception block may use this information when identifying the object. In other instances, the secondary computer may have its own neural network that is trained as described herein, thus reducing the risk of false positives.

[0197] The mobile vehicle 1300 may further include an infotainment SoC 1330 (e.g., an in-vehicle infotainment system (IVI)). Although shown and described as an SoC, the infotainment system may not be an SoC and may include two or more separate components. The infotainment SoC 1330 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., TV, movies, streaming, etc.), telephony (e.g., hands-free calling), network connectivity (e.g., LTE, Wi-Fi, etc.), and / or information services (e.g., navigation systems, rear parking assist, wireless data systems, mobile vehicle related information such as fuel level, total distance traveled, brake fuel level, oil level, door opening / closing, air filter information, etc.). For example, the infotainment SoC 1330 may be a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, a car computer, in-car entertainment, Wi-Fi, steering wheel audio controls, hands-free voice control, a heads-up display (HUD), an HMI display 1334, 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 SoC 1330 may further be used to provide information (e.g., visual and / or audible) to a user of the vehicle, such as information from an ADAS system 1338, autonomous driving information such as planned vehicle maneuvers, trajectory, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0198] The infotainment SoC 1330 may include GPU functionality. The infotainment SoC 1330 may communicate with other devices, systems, and / or components of the mobile vehicle 1300 via the bus 1302 (e.g., CAN bus, Ethernet, etc.). In some instances, the infotainment SoC 1330 may be coupled to a supervisory MCU such that the infotainment system's GPU can perform some self-drive functions in the event that the primary controller 1336 (e.g., the primary and / or backup computer of the mobile vehicle 1300) fails. In such instances, the infotainment SoC 1330 can place the mobile vehicle 1300 in a Chauffeur safe stop mode, as described herein.

[0199] The mobile vehicle 1300 may further include an instrument cluster 1332 (e.g., a digital dash, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 1332 may include a controller and / or a supercomputer (e.g., a separate controller or a supercomputer). The instrument cluster 1332 may include a set of instruments such as a speedometer, fuel level, oil pressure, a tachometer, an odometer, turn signals, a gear shift position indicator, a seat belt warning light, a parking brake warning light, an engine malfunction light, an airbag (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some instances, information may be displayed and / or shared between the infotainment SoC 1330 and the instrument cluster 1332. In other words, the instrument cluster 1332 may be included as part of the infotainment SoC 1330, or vice versa.

[0200] 13D is a system diagram of communication between a cloud-based server and an example autonomous vehicle 1300 of FIG. 13A in accordance with some embodiments of the present disclosure. System 1376 may include a server 1378, a network 1390, and a mobile vehicle including mobile vehicle 1300. Server 1378 may include multiple GPUs 1384(A)-1384(H) (collectively referred to herein as GPUs 1384), PCIe switches 1382(A)-1382(H) (collectively referred to herein as PCIe switches 1382), and / or CPUs 1380(A)-1380(B) (collectively referred to herein as CPUs 1380). GPUs 1384, CPUs 1380, and PCIe switches may be interconnected with a high-speed interconnect, such as, but not limited to, an NVLINK interface 1388 and / or a PCIe connection 1386 developed by NVIDIA. In some instances, the GPUs 1384 are connected via an NVLINK and / or NVSwitch SoC, and the GPUs 1384 and PCIe switch 1382 are connected via a PCIe interconnect. Although eight GPUs 1384, two CPUs 1380, and two PCIe switches are illustrated, this is not intended to be limiting. Depending on the embodiment, each server 1378 may include any number of GPUs 1384, CPUs 1380, and / or PCIe switches. For example, the servers 1378 may each include 8, 16, 32, and / or more GPUs 1384.

[0201] The server 1378 can receive image data from the mobile vehicle over the network 1390, the image data representing images showing unexpected or changed road conditions, such as recently started road construction. The server 1378 can transmit the neural network 1392, the updated neural network 1392, and / or the map information 1394, including information about traffic and road conditions, to the mobile vehicle over the network 1390. The updates to the map information 1394 can include updates to the HD map 1322, such as information about construction sites, potholes, detours, flooding, and / or other obstacles. In some instances, the neural network 1392, the updated neural network 1392, and / or the map information 1394 may have arisen from new training and / or experience represented in data received from any number of mobile vehicles in the environment and / or based on training performed at a data center (e.g., using the server 1378 and / or other servers).

[0202] The server 1378 may be used to train a machine learning model (e.g., a neural network) based on the training data. The training data may be generated by the vehicle and / or generated in a simulation (e.g., using a game engine). In some instances, the training data is tagged (e.g., if the neural network benefits from supervised learning) and / or undergoes other pre-processing, while in other instances, the training data is not tagged and / or pre-processed (e.g., if the neural network does not require supervised learning). After the machine learning model is traced, it may be used by the vehicle (e.g., transmitted to the vehicle via the network 1390) and / or used by the server 1378 to remotely monitor the vehicle.

[0203] In some instances, the server 1378 can receive data from the moving vehicles and apply the data to state-of-the-art real-time neural networks for real-time intelligent inference. The server 1378 can include deep learning supercomputers and / or dedicated AI computers powered by the GPU 1384, such as the DGX and DGX Station machines developed by NVIDIA. However, in some instances, the server 1378 can include a deep learning infrastructure that uses only CPU-powered data centers.

[0204] The deep learning infrastructure of the server 1378 may be capable of rapid real-time inference and may use that capability to evaluate and validate the health of the processor, software, and / or associated hardware within the moving vehicle 1300. For example, the deep learning infrastructure may receive periodic updates from the moving vehicle 1300, such as sequences of images and / or objects that the moving vehicle 1300 has located within the images of that sequence (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 with objects identified by the moving vehicle 1300, and if the results do not match and the infrastructure concludes that the AI ​​within the moving vehicle 1300 is not functioning properly, the server 1378 may send a signal to the moving vehicle 1300 to infer control, notify passengers, and command the failsafe computer of the moving vehicle 1300 to complete a safe parking maneuver.

[0205] For inference, the server 1378 may include a GPU 1384 and one or more programmable inference accelerators (e.g., NVIDIA's TensorRT). The combination of a GPU-powered server and inference acceleration can enable real-time responsiveness. In other instances, such as when less performance is required, servers powered by CPUs, FPGAs, and other processors may be used for inference.

[0206] Exemplary Computing Device 14 is a block diagram of an example computing device 1400 suitable for use in implementing some embodiments of the present disclosure. The computing device 1400 may include a bus 1402 that indirectly or directly couples the following devices: memory 1404, one or more central processing units (CPUs) 1406, one or more graphic processing units (GPUs) 1408, a communication interface 1410, input / output (I / O) ports 1412, input / output components 1414, a power supply 1416, and one or more presentation components 1418 (e.g., a display).

[0207] Although the various blocks in FIG. 14 are shown connected by lines via bus 1402, this is not intended to be limiting and is merely for clarity. For example, in some embodiments, a presentation component 1418, such as a display device, may be considered an I / O component 1414 (e.g., if the display is a touch screen). As another example, CPU 1406 and / or GPU 1408 may include memory (e.g., memory 1404 may represent a storage device in addition to the memory of GPU 1408, CPU 1406, and / or other components). In other words, the computing devices of FIG. 14 are merely exemplary. Categories such as "workstation," "server," "laptop," "desktop," "tablet," "client device," "mobile device," "handheld device," "gaming machine," "electronic control unit (ECU)," "virtual reality system," and / or other device or system types are all intended to be within the scope of the computing devices of FIG. 14 and therefore are not to be differentiated.

[0208] Bus 1402 may represent one or more buses, such as an address bus, a data bus, a control bus, or a combination thereof. Bus 1402 may include one or more bus types, such as an industrial standard architecture (ISA) bus, an extended industrial standard architecture (EISA) bus, a video electronics standard association (VESA) bus, a peripheral component interconnect (PCI) bus, a PCI express (PCIe) bus, and / or another type of bus.

[0209] Memory 1404 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 1400. Computer readable media may include both volatile and nonvolatile media, and removable and non-removable media. By way of example, and not limitation, computer readable media may comprise computer storage media and communication media.

[0210] Computer storage media may include both 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 1404 may store computer readable instructions (e.g., representing programs and / or program elements), such as an operating system. Computer storage media may include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that may be used to store the desired information and that may be accessed by computing device 1400. As used herein, computer storage media does not include the signals themselves.

[0211] 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 such a manner as to encode 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. Combinations of any of the above should also be included within the scope of computer-readable media.

[0212] The CPU 1406 may be configured to execute computer readable instructions to control one or more components of the computing device 1400 to perform one or more of the methods and / or processes described herein. The CPU 1406 may include one or more (e.g., 1, 2, 4, 8, 28, 72, etc.) cores each capable of processing multiple software threads simultaneously. The CPU 1406 may include any type of processor, and may include different types of processors depending on the type of computing device 1400 implemented (e.g., a processor with fewer cores for a mobile device and a processor with more cores for a server). For example, depending on the type of computing device 1400, 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). The computing device 1400 may include one or more CPUs 1406 within one or more microprocessors or auxiliary coprocessors, such as computational coprocessors.

[0213] The GPU 1408 may be used by the computing device 1400 to render graphics (e.g., 3D graphics). The GPU 1408 may include hundreds or thousands of cores capable of processing hundreds or thousands of software threads simultaneously. The GPU 1408 may generate pixel data for an output image in response to a rendering command (e.g., a rendering command from the CPU 1406 received via a host interface). The GPU 1408 may include graphics memory, such as a display memory, for storing the pixel data. The display memory may be included as part of the memory 1404. The GPU 1408 may include two or more GPUs operating in parallel (e.g., via a link). When combined, each GPU 1408 may generate pixel data for a different portion 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 the other GPUs.

[0214] In instances where the computing device 1400 does not include a GPU 1408, the CPU 1406 may be used to render graphics.

[0215] The communications interface 1410 may include one or more receivers, transmitters, and / or transceivers that enable the computing device 700 to communicate with other computing devices over electronic communications networks, including wired and / or wireless communications. The communications interface 1410 may include components and functionality to enable 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.

[0216] The I / O ports 1412 may enable the computing device 1400 to be logically coupled to other devices, including I / O components 1414, presentation components 1418, and / or other components, some of which may be built-in (e.g., integrated) to the computing device 1400. Exemplary I / O components 1414 include a microphone, a mouse, a keyboard, a joystick, a game pad, a game controller, a satellite dish, a scanner, a printer, a wireless device, and the like. The I / O components 1414 may provide a natural user interface (NUI) that processes air gestures, voice, 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 voice recognition, stylus recognition, facial recognition, biometric recognition, gesture recognition on and next to the screen, air gestures, head and eye tracking, and touch recognition in conjunction with the display of the computing device 1400 (as described in more detail below). The computing device 1400 may include a depth camera, such as a stereoscopic camera system, an infrared camera system, an RGB camera system, touch screen technology, and combinations thereof, for gesture detection and recognition. Additionally, the computing device 1400 may include an accelerometer or gyroscope (e.g., as part of an inertia measurement unit (IMU)) to enable detection of motion. In some instances, the output of the accelerometer or gyroscope may be used by the computing device 1400 to render immersive augmented or virtual reality.

[0217] The power supply 1416 may include a hardwired power supply, a battery power supply, or a combination thereof. The power supply 1416 may provide power to the computing device 1400 to enable the components of the computing device 1400 to operate.

[0218] The presentation component 1418 may include a display (e.g., a monitor, a touch screen, a television screen, a heads-up display (HUD), other display types, or a combination thereof), speakers, and / or other presentation components. The presentation component 1418 can receive data from other components (e.g., GPU 1408, CPU 1406, etc.) and output data (e.g., as images, video, sound, etc.).

[0219] 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, being executed by a computer or other machine, such as a personal digital assistant or other handheld device. Generally, program modules, including routines, programs, objects, components, data structures, etc., refer to code that performs particular tasks or implements particular abstract data types. The present disclosure may be implemented in a variety of configurations, including handheld devices, consumer electronics, general purpose computers, more specialized computing devices, etc. The present disclosure may also be implemented in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network.

[0220] As used herein, the term "and / or" with respect to two or more elements should be construed to mean only one element or a combination of elements. For example, "element A, element B, and / or element C" may include element A only, element B only, element C only, elements A and element B, elements A and element C, elements 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. Furthermore, "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.

[0221] The subject matter of the present disclosure has been described with specificity to meet statutory requirements. However, the description itself is not intended to limit the scope of the present disclosure. Instead, the inventors intend that the claimed subject matter may be implemented in other ways, including different steps or combinations of steps similar to those described in this document, in conjunction with other current or future technologies. Furthermore, although the terms "step" and / or "block" may be used herein to connote different elements of the method used, these terms should not be construed as implying any particular order between the various steps disclosed herein unless and when the order of the individual steps is explicitly described.

Claims

1. receiving RADAR and LIDAR data collected from a three dimensional (3D) environment; orthogonally projecting a first instance of the LIDAR data associated with a time slice to generate a LIDAR projection image representative of LIDAR detections within the 3D environment; receiving a LIDAR label identifying a location of an object represented by the LIDAR detection in the LIDAR projection image; orthogonally projecting a second instance of the RADAR data associated with the time slice to generate a RADAR projection image representative of RADAR detections within the 3D environment; imparting the LIDAR label to the RADAR projection image to generate a propagated LIDAR label; determining a number of the RADAR detections corresponding to each of the transmitted LIDAR labels; removing the set of propagated LIDAR labels that include fewer than a threshold number of RADAR detections, leaving a set of remaining LIDAR labels; training a neural network to detect objects in the RADAR projection images using ground truth data generated based at least in part on the set of remaining LIDAR labels; A method comprising:

2. encoding the RADAR projection images and a set of features representative of the RADAR detections and corresponding reflectance properties into a multi-channel RADAR data tensor; training the neural network using the multi-channel RADAR data tensor and the ground truth data as training data; The method of claim 1 further comprising:

3. The method of claim 1 , wherein the second instance of the RADAR data comprises accumulated ego-motion corrected RADAR detections.

4. The method of claim 1 , wherein the ground truth data includes a class confidence tensor and an instance regression tensor.

5. 2. The method of claim 1 , further comprising: generating the ground truth data using the remaining set of LIDAR labels to generate at least one of position, size, or orientation data for the objects represented by the LIDAR detections; and encoding the at least one of position, size, or orientation data into a corresponding channel of an instance regression tensor.

6. 2. The method of claim 1 , wherein the LIDAR labels further identify a class of the object represented by the LIDAR detection in the LIDAR projection image, and the method further comprises generating the ground truth data by encoding classification data representative of the class of the object into a corresponding channel of a class confidence tensor.

7. 2. The method of claim 1 , wherein the LIDAR labels include bounding boxes drawn around stationary vehicles in the 3D environment, the method further comprising training the neural network to detect other stationary vehicles from input RADAR data using the bounding boxes drawn around the stationary vehicles.

8. 2. The method of claim 1, wherein the LIDAR projection image is an orthogonal projection of a LIDAR point cloud, and the LIDAR labels include a set of closed polylines around each vehicle drawn on top of the orthogonal projection.

9. receiving RADAR and LIDAR data collected from a three dimensional (3D) environment; For each frame of the LIDAR data, generating a LIDAR projection image from the frames of LIDAR data representative of LIDAR detections from the 3D environment; receiving a LIDAR label identifying a location of an object represented by the LIDAR detection in the LIDAR projection image; generating a RADAR projection image representative of RADAR detections from the 3D environment from a corresponding frame of the RADAR data associated with the frame of LIDAR data; training a neural network to detect objects in the RADAR projection images using ground truth data generated based at least in part on the LIDAR labels; Including, imparting the LIDAR label to the RADAR projection image to generate a propagated LIDAR label; determining a number of the RADAR detections corresponding to each of the transmitted LIDAR labels; and removing the set of propagated LIDAR labels that include fewer than a threshold number of RADAR detections, leaving a set of remaining LIDAR labels; generating the ground truth data from the remaining set of LIDAR labels; generating the ground truth data by

10. encoding the RADAR projection images and a set of features representative of the RADAR detections and corresponding reflectance properties into a multi-channel RADAR data tensor; training the neural network using the multi-channel RADAR data tensor and the ground truth data as training data; The method of claim 9 further comprising:

11. The method of claim 9 , wherein the corresponding frames of RADAR data include accumulated ego-motion corrected RADAR detections.

12. The method of claim 9 , wherein the ground truth data includes class confidence tensors and instance regression tensors.

13. 10. The method of claim 9, further comprising: generating position, size, and orientation data for the objects represented by the LIDAR detections using the set of LIDAR labels to generate the ground truth data; and encoding the position, size, and orientation data into corresponding channels of an instance regression tensor.

14. 10. The method of claim 9, wherein the LIDAR labels further identify a class of the object represented by the LIDAR detection in the LIDAR projection image, and the method further comprises generating the ground truth data by encoding classification data representative of the class of the object into a corresponding channel of a class confidence tensor.

15. receiving a LIDAR label of a LIDAR detection in one or more instances of LIDAR data collected from an environment over a period of time using a LIDAR sensor; For each of the one or more instances of LIDAR data, identifying a closest instance of RADAR data to said instance of LIDAR data, said closest instance of RADAR data being associated with an accumulated, ego-motion corrected, orthogonal image of a portion of said RADAR data collected from said environment over said period of time using a RADAR sensor; communicating the set of LIDAR labels for the LIDAR detections in the instance of LIDAR data to a corresponding set of RADAR detections in the closest instance of RADAR data; A method comprising: transmitting the set of LIDAR labels to form transmitted LIDAR labels; The transmitting step further comprises: removing the set of propagated LIDAR labels that include fewer than a threshold number of the RADAR detections, leaving a set of remaining LIDAR labels; training a neural network to detect objects in the RADAR data using ground truth data generated from the remaining set of LIDAR labels; The method further comprising:

16. The method of claim 15 , further comprising training a neural network to detect objects in the RADAR data using ground truth data generated from the set of LIDAR labels.

17. 16. The method of claim 15, wherein the LIDAR labels include bounding boxes drawn around stationary vehicles in a 3D environment, the method further comprising training a neural network to detect other stationary vehicles from input RADAR data using the bounding boxes drawn around the stationary vehicles.

18. encoding the nearest instances of RADAR data and a set of features representative of corresponding RADAR detection and reflectance characteristics into a multi-channel RADAR data tensor; training a neural network to detect objects in the RADAR data using the multi-channel RADAR data tensor and ground truth data generated from the set of LIDAR labels; The method of claim 15 further comprising:

Images