Methods and systems for sensor fusion for autonomous machine applications

Abstract

In various examples, a multi-sensor fusion machine learning model, such as a deep neural network (DNN), can be deployed to fuse data from multiple individual machine learning models. Thus, the multi-sensor fusion network can use the outputs from multiple machine learning models as input to generate a fused output that represents data from the field of view or sensory field of each sensor feeding the machine learning model, while taking into account the learned associations between the boundaries or overlapping regions of the respective fields of view of the source sensors. In this way, the fused output may be less likely to include duplicate, inaccurate, or noisy data related to objects or features in the environment because the fusion network can be trained to interpret multiple instances of the same object appearing in different input representations.

Description

Background Art

[0001] The ability to safely detect static and dynamic objects, obstacles, hazards, waiting conditions, road markings, signs, and / or other features of the environment is a critical task for any autonomous or semi-autonomous driving system. For example, detecting the location of various static and / or dynamic features or objects in an environment in three-dimensional (3D) space is particularly challenging and can result in noisy or inaccurate data—particularly with respect to boundaries or overlapping areas between different sensor fields of view.

[0002] In some systems, deep neural networks (DNNs) can be deployed to generate 3D information corresponding to the field of view of an associated sensor that provides sensor data to the DNN. For example, for objects or landmarks that appear near the boundaries of an image, these perception systems may result in noisy results—for example, resulting in rapid changes between detecting and not detecting the same object or landmark—when the object or landmark disappears and reappears in the sensor's field of view. As a result, for example, when tracking objects or landmarks across frames, the results may be inaccurate or imprecise, and may lead to degraded system performance.

[0003] To address these potential shortcomings of individual DNNs, some conventional systems employ hand-crafted or rule-based algorithms to stitch together the outputs from multiple DNNs. However, in situations where the fields of view between adjacent sensors providing sensor data to individual DNNs overlap, fused predictions along the boundaries or overlapping regions may be inconsistent or inaccurate. For example, a DNN operating on a sensor with a 120-degree field of view may predict that an object 10 meters from the sensor is 9 meters away, while another DNN operating on a sensor with a 60-degree field of view may predict that the object is 11 meters away. In such an example, even though the actual distances of the individual DNNs to the object differ by only one meter, the outputs differ by two meters, which may be a large enough distance gap to cause the fused output to include two separate, duplicate objects rather than the single object actually represented in the data. The underlying system may then rely on this erroneous determination that two separate objects exist, and the false detection can propagate through object tracking, planning, control, obstacle avoidance, and / or other operations of the machine. Summary of the Invention

[0004] Embodiments of the present disclosure relate to machine learning-based sensor fusion for autonomous machine applications. Systems and methods are disclosed for fusing data from multiple separate machine learning models using a multi-sensor fusion machine learning model (e.g., a deep neural network (DNN)). For example, multiple machine learning models can be trained to generate outputs that can be used as inputs to a multi-sensor fusion network, either directly or after post-processing. Each output received as input by the multi-sensor fusion network may correspond to the same type of representation—e.g., a rasterized image from the same perspective (e.g., a top-down, egocentric perspective). Thus, the multi-sensor fusion network can use these inputs to generate a fused output that represents data from the field of view or sensory field of each sensor type, while taking into account learned associations between the boundaries or overlapping regions of the respective fields of view of the source sensors. In this way, the fused output may be less likely to include duplicate, inaccurate, or noisy data related to objects or features in the environment because the fusion network can be trained to interpret multiple instances of the same object appearing in different inputs.

[0005] In some embodiments, in addition to or instead of using the outputs from the various source machine learning models, internal values of the various machine learning models—for example, output values of one or more feature extractor layers from the various DNNs—can be provided as inputs to the multi-sensor fusion network. In such an example, the multi-sensor fusion network and selected layers of the various machine learning models can be trained together in an end-to-end training process. Thus, updates to weights and biases due to one or more loss functions can be back-propagated not only through the layers of the multi-sensor fusion network, but also through the layers of the corresponding source machine learning models (for example, feature extraction layers).

[0006] To further improve the accuracy and precision of the multi-sensor fusion network, one or more additional channels can be provided as inputs to the multi-sensor fusion network—e.g., at each iteration. For example, a location prior channel can be used to represent the uncertainty of the distance estimate—e.g., as a probability distribution function (PDF)—at one or more pixels or points corresponding to the field of view or sensor field of one or more source sensors. In some embodiments, a velocity image channel can be used as input to the multi-sensor fusion network to represent the associated velocities of objects detected in the environment. For example, different sensors or machine learning model outputs can include velocity information—e.g., in the x and / or y directions—and the associated velocities of the objects can be used to fuse objects together (e.g., where the velocity predictions closely match) or to represent two or more objects (e.g., where the velocity predictions for closely located objects differ by more than a threshold). Additionally, instance- and / or appearance-based spatial and / or temporal correlation channels can be used as additional inputs to the multi-sensor fusion network to help identify similar object appearances across different inputs and / or similar object appearances across time steps—e.g., by tracking objects across time using a recurrent neural network.

[0007] As a result, the underlying machine learning model that feeds the multi-sensor fusion network can be optimized to generate outputs that are useful for improving the accuracy of the fusion output of the multi-sensor fusion network. Additional inputs—e.g., position prior channels, velocity image channels, etc.—can be used to further improve the accuracy and precision of the multi-sensor fusion network, particularly at the boundaries or overlapping regions between adjacent sensor fields of view or sensory fields. By reducing noise and improving the accuracy and precision of the multi-sensor fusion network, downstream processes that rely on these outputs of the multi-sensor fusion network—representing the position, velocity, pose, appearance, etc. of static and dynamic objects or features in the environment—can also benefit from the improvements in performance and effectiveness. BRIEF DESCRIPTION OF THE DRAWINGS

[0008] The present system and method for machine learning-based sensor fusion for autonomous machine applications are described in detail below with reference to the accompanying drawings, wherein:

[0009] Figure 1A and Figure 1B is an example data flow diagram corresponding to a process and for multi-sensor fusion according to some embodiments of the present disclosure;

[0010] Figures 2A-2E is an example visualization of the input of a multi-sensor fusion network according to some embodiments of the present disclosure;

[0011] Figures 3A-3B is an example visualization of the output of a multi-sensor fusion network according to some embodiments of the present disclosure;

[0012] Figures 4A-4C is an example visualization of the fields of view or sensor fields of various sensors of an ego machine and associated rays from the various sensors intersecting perceived objects in an environment, according to some embodiments of the present disclosure;

[0013] Figure 5 is a flow chart illustrating a method of using a multi-sensor fusion network to compute a fused output using multiple input channels according to some embodiments of the present disclosure;

[0014] Figures 6A-6B A data flow diagram representing a process for training a multi-sensor fusion network and one or more associated source networks according to some embodiments of the present disclosure;

[0015] Figure 6C is an example of ground truth data for training a multi-sensor fusion network according to some embodiments of the present disclosure;

[0016] Figure 7 is a flowchart illustrating a method for training a multi-sensor fusion network to compute a fusion output using multiple input channels according to some embodiments of the present disclosure;

[0017] Figure 8A is an illustration of an example autonomous vehicle according to some embodiments of the present disclosure;

[0018] Figure 8B According to some embodiments of the present disclosure Figure 8A Examples of camera positions and fields of view for autonomous vehicles;

[0019] Figure 8C According to some embodiments of the present disclosure Figure 8A a block diagram of an example system architecture for an example autonomous vehicle;

[0020] Figure 8D According to some embodiments of the present disclosure, a method for Figure 8A System diagram of an example of communication between autonomous vehicles;

[0021] Figure 9 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure; and

[0022] Figure 10 is a block diagram of an example data center suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0023] Systems and methods are disclosed related to machine learning-based sensor fusion for autonomous machine applications. Although the present disclosure may be directed to an example autonomous vehicle 800 (alternatively referred to herein as "vehicle 800" or "autonomous vehicle 800" or "ego vehicle 800"), its examples are directed to Figures 8A-8D The present disclosure may be described with respect to sensor fusion for autonomous or semi-autonomous machine operation, but this is not intended to be limiting. For example, the systems and methods described herein may be used with, but not limited to, non-autonomous vehicles, semi-autonomous vehicles (e.g., in one or more adaptive driver assistance systems (ADAS)), manned and unmanned robots or robotic platforms, warehouse vehicles, off-road vehicles, vehicles connected to one or more trailers, flying boats, boats, shuttles, emergency vehicles, motorcycles, electric or motorized bicycles, aircraft, engineering vehicles, underwater vehicles, drones, and / or other vehicle or machine types. Furthermore, while the present disclosure may be described with respect to sensor fusion for autonomous or semi-autonomous machine operation, this is not intended to be limiting, and the systems and methods described herein may be used in augmented reality, virtual reality, mixed reality, robotics, security and surveillance, autonomous or semi-autonomous machine applications, and / or any other technology space where sensor fusion may be used.

[0024] See also Figures 1A-1B , Figures 1A-1B is an example data flow diagram corresponding to processes 100A and 100B (collectively referred to herein as "process 100") for multi-sensor fusion according to some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are set forth as examples only. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, functional groupings, etc.) may be used in addition to or in place of those shown, and some elements may be omitted altogether. Further, many of the elements described herein are functional entities that may be implemented as discrete or distributed components, or in combination with other components and in any suitable combination and location. The various functions described herein as being performed by entities may be performed by hardware, firmware, and / or software. For example, the various functions may be performed by a processor executing instructions stored in a memory. In some embodiments, processes 100A and 100B may be performed using Figures 8A-8D components, features and / or functions of the machine 800, Figure 9 The example computing device 900 and / or Figure 10 The example data center 1000 is used for execution.

[0025] Process 100 may include generating, accessing, and / or receiving sensor data from one or more sensors. As a non-limiting example, sensor data may be obtained from a vehicle (e.g., as described herein). Figures 8A-8DOne or more sensors of the vehicle 800 receive sensor data. The sensor data can be used by the vehicle 800 and within the process 100 to use a corresponding DNN corresponding to the camera sensor, a 3D RADAR signal (S RADAR ) 108 (which can be calculated using RADAR data alone and / or can be calculated using DNN), 3D ultrasonic signals (S USS) 112 (which can be calculated using ultrasonic data alone and / or can be calculated using a DNN), a 3D LiDAR signal (S LiDAR )(not shown) (which can be calculated using LiDAR data alone and / or can be calculated using a DNN) to calculate two-dimensional (2D) and / or three-dimensional (3D) signals (S1-S n ) 104A-104N, and / or for using additional or alternative sensor data (from any sensor modality) and / or corresponding DNNs to compute other output signals. In some embodiments, sensor data can be used to compute outputs within the DNN itself, such as feature outputs (F1-F2) of a DNN using camera signals. n )102A-102N, RADAR sensor RADAR feature output (F RADAR )106、Ultrasonic characteristic output of ultrasonic sensor (F USS )110, LiDAR feature output of LiDAR sensor (F LiDAR ) and / or other feature outputs of other sensor types. Thus, the sensor data can be computed using any number of sensors and using any number of different sensor modalities, and the sensor data can be used directly (e.g., with or without preprocessing, such as converting raw LiDAR data into a point cloud) and / or can be used after being processed by a DNN or other type of machine learning model.

[0026] During training (this paper focuses on Figures 6A-6C 、 Figure 7), sensor data may be generated using one or more data collection vehicles that generate sensor data for training a DNN (such as, for example, DNN(s) associated with feature outputs and / or 3D signals), and / or may be pre-generated and included in a training dataset. The sensor data used in the training process may additionally or alternatively be generated using simulated sensor data (e.g., sensor data generated using one or more virtual sensors of a virtual vehicle in a virtual environment) and / or augmented sensor data (e.g., sensor data generated using one or more data collection vehicles and modified using virtual data, respectively). Once trained and deployed in vehicle 800, the sensor data may be generated by one or more sensors of vehicle 800 and processed by DNN(s) to compute various output signals and / or compute feature outputs from one or more feature extractor layers of each DNN.

[0027] For non-limiting example, sensor data may include data from any sensor of the vehicle 800, such as and referring to Figures 8A-8C , any sensor of the vehicle 800 includes (one or more) RADAR sensors 860, (one or more) ultrasonic sensors 862, one or more LIDAR sensors 864, one or more stereo cameras 868, one or more wide-angle cameras 870 (e.g., fisheye cameras), (one or more) infrared cameras 872, (one or more) surround cameras 874 (e.g., 360-degree cameras), (one or more) long-range and / or medium-range cameras 878, and / or other sensor types. As another example, the sensor data may include virtual (e.g., simulated or augmented) sensor data generated from any number of sensors of a virtual vehicle or other virtual object in a virtual (e.g., test) environment. In such an example, the virtual sensors may correspond to virtual vehicles or other virtual objects in a simulated environment (e.g., for testing, training, and / or validating DNN performance), and the virtual sensor data may represent sensor data captured by the virtual sensors within the simulated or virtual environment. Thus, by using virtual sensor data, the DNN(s) described herein may be tested, trained, and / or validated in a simulated environment using simulated or augmented data, which may allow for testing of more extreme scenarios outside of real-world environments where such testing may be less safe.

[0028] In some embodiments, the sensor data may include image data representing (one or more) images, image data representing a video (e.g., a snapshot of a video), and / or sensor data representing a representation of a sensor's sensing field (e.g., a depth map, a point cloud, a range image, etc. for a LiDAR sensor, a value map for an ultrasonic sensor, etc.). Where the sensor data includes image data, any type of image data format may be used, such as, but not limited to, compressed images (e.g., Joint Photographic Experts Group (JPEG) or Luma / Chroma (YUV) formats), compressed images as frames derived from a compressed video format (e.g., H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC)), such as raw images derived from red-clear-blue (RCCB), red-clear-color (RCCC), or other types of imaging sensors, and / or other formats. Additionally, in some examples, sensor data may be used within process 100 without any preprocessing (e.g., in raw or captured format), while in other examples, the sensor data may undergo preprocessing (e.g., noise balancing, demosaicing, scaling, cropping, enhancement, white balancing, tone curve adjustment, point cloud generation, projected image generation (e.g., to generate a LiDAR range image), etc., such as using a sensor data preprocessor (not shown)). As used herein, sensor data may refer to unprocessed sensor data, preprocessed sensor data, or a combination thereof.

[0029] In the example where a LiDAR sensor is used to generate, for example, a point cloud, point clouds from one or more LiDAR sensors can be merged together—for example, after performing motion compensation for the ego machine 800 and / or performing data synchronization to address motion-related LiDAR and time synchronization issues. Similarly, combinations of sensor data representations (e.g., point clouds, depth maps, etc.) of different sensor types (e.g., RADAR, ultrasonic, etc.) can also undergo similar pre-processing before being used directly and / or processed by one or more machine learning models to compute a 3D signal.

[0030] For training, as described in more detail herein, sensor data may include raw images (e.g., as captured by one or more image sensors), downsampled images, upsampled images, cropped or region of interest (ROI) images, otherwise enhanced images, and / or combinations thereof. Similarly, where the sensor data corresponds to data other than image data, the sensor data used for training may include different orientations, projections, reference points, enhancements, cropping, filtering, etc. During training of one or more DNNs (including fusion DNN 120), ground truth data may be generated. In some embodiments, the ground truth data may be automatically generated in an unsupervised manner (e.g., using a photometric consistency loss between outputs of sensors with at least partially overlapping fields of view or sensory fields), may be generated in a supervised manner, for example, using annotation data, and / or may be generated in a semi-supervised manner. Where annotations are used to generate the ground truth data, the annotations may be generated within a drawing program (e.g., an annotation program), a computer-aided design (CAD) program, a labeling program, another type of program suitable for generating annotations, and / or, in some examples, may be manually drawn. In any example, the annotation data can be synthetically generated (e.g., generated from a computer model or rendering), realistically generated (e.g., designed and generated from real-world data), machine-automated (e.g., using feature analysis and learning to extract features from the data and then generate labels), human-annotated (e.g., a labeler or annotation expert defines the location of the labels), and / or a combination thereof (e.g., a human identifies the center or origin and dimensions of the region, and a machine generates polygons and / or labels for the intersection region).

[0031] In an embodiment, each feature output or 3D signal may correspond to a respective sensor pipeline or stream. For example, a first sensor pipeline may include a first camera that may generate image data that may be processed by a first DNN to generate feature output F1 and / or 3D signal 104A, a second sensor pipeline may include a second camera that may generate image data that may be processed by a second DNN to generate feature output F2 and / or 3D signal 104B, and a third sensor pipeline may include a first RADAR sensor that may generate RADAR data that may be directly processed—e.g., using a sensor data preprocessor—and / or may be processed using a DNN to generate feature output F2 and / or 3D signal 104B. RADAR and / or 3D signal 108 , etc. Depending on the embodiment, any number of sensor lines may be used.

[0032] Where process 100 utilizes a 3D signal, the 3D signal may be computed by each respective DNN in the same format, or may be converted—e.g., using a post-processor—into the same format. For example, in some non-limiting embodiments, the 3D signal used as input to fused DNN 120 may include a rasterized image generated from a certain perspective (e.g., a top-down bird's-eye view, a projection image such as a range image, a side view image, etc.) encoding any number of input channels. In embodiments where the ego machine 800 is located at the center of the representation, the rasterized image—including the fused output rasterized image computed using the fused DNN—may be egocentric. In other embodiments, the input 3D signal may be generated from the perspective of the ego machine 800, and the fused output 122 may be generated from an egocentric perspective. For example, the input channels may indicate the shape, orientation, and / or classification of objects or features in the environment. In such an example, the rasterized image may include, for example, a bounding shape or cuboid corresponding to a dynamic actor, and may include lines corresponding to lane markings (e.g., lane dividers, road dividers, solid lines, dashed lines, double lines, yellow lines, white lines, etc.), waiting conditions (e.g., crosswalks, stop lines, etc.) and / or other driving surface features, boundaries or encoded values of pixels corresponding to drivable free space (e.g., an area of the environment that the self-machine 800 can traverse), and / or other objects or features.

[0033] Input channels included in the rasterized input image (or other input representation) may include, but are not limited to, object or feature position or occupancy channels, which may include a starting height of an object as a channel and / or an ending height of an object as a channel (e.g., where a top-down view rasterized image is generated, a pixel position may represent an x or y pixel position of a corresponding feature or object, indicating a horizontal or vertical position relative to the ego machine 800, and the pixel may be encoded with one or more height channels indicating the height of the object or feature), one or more velocity channels (e.g., a pixel of an image may be encoded with velocity in the x and / or y directions), a direction channel corresponding to one or more objects (e.g., encoded as one or more angles), and / or a classification channel corresponding to one or more objects or features, and / or additional or alternative channels. Furthermore, these channels may be rasterized for each different input image corresponding to a corresponding sensor pipeline. For example, for a camera sensor pipeline including 3D signal 104A, there may be (one or more) occupancy channels, (one or more) velocity channels, (one or more) classification channels, etc., and for a RADAR sensor pipeline including 3D signal 108, there may be (one or more) occupancy channels, (one or more) velocity channels, (one or more) classification channels, etc. for each sensor pipeline input representation, and so on. In some embodiments, different input signals to the fused DNN 120 may include different channel types, the same channel type, or a combination thereof. For non-limiting example, the first sensor pipeline may generate 3D signal 104A including (one or more) occupancy channels and (one or more) velocity channels, while the 3D signal 112 may only include (one or more) occupancy channels.

[0034] Non-limiting examples of input channels of different sensor types that may be included in an input 2D or 3D signal are Figure 1B , where various input signals 104, 108, 112, 114, 116, and / or 118 may include rasterized images representing one or more of the input channel types shown. Other examples of 3D signals, such as where a 3D signal includes a rasterized image, are shown. Figure 2A and Figure 2B For example, Figure 2AThe image may represent: a rasterized image 202A from a first sensor pipeline including sensors with a field of view or sensory field in front of the ego machine 800; and a second rasterized image 202B from a second sensor pipeline including sensors with a field of view or sensory field behind the ego machine 800. Rasterized images 202A and 202B may include lane markings 204A, 204B, etc., object detections 206A, 206B, 206C, etc. (e.g., corresponding to vehicles, pedestrians, animals, cyclists, debris, robots, etc.), and / or indications of free space (e.g., pixels may be encoded as free space or non-free space). As shown with respect to detections 206, the detections may include a bounding shape indicating the location, shape, and / or orientation of the detected objects (e.g., points in the bounding shape indicate a direction of travel or directional position). Object detections 206, lane markings 204, and / or other features or objects encoded in rasterized image 202 may also include classification information, speed information, and / or other information. As described in this article, in addition to Figures 2A-2B Additional or alternative features or objects beyond those shown in may be included in the rasterized image 202, such as waiting conditions, road contour information (eg, potholes, disturbances, speed bumps, etc.), and / or other information describing the surrounding environment.

[0035] In some embodiments, the classification information corresponding to an object or feature can be represented by an intensity value. For example, if a classification corresponds to an intensity value of x, then the overlapping area of two objects will be x+1, the overlapping area of three objects will be x+2, and so on. However, in the case of overlap among objects, the boundaries of the objects can be encoded with an intensity value indicating the boundary, such as, but not limited to, a maximum intensity value of 255, where the intensity value range is 0-255. As a result, the boundaries between detected objects can be more easily identified.

[0036] Different input 3D signals may correspond to different sensors with different fields of view or perception. Figures 4A-4C as well as Figure 8B As shown, various sensor configurations may be used to compute 3D signals corresponding to some or all of the environment of the self-machine 800. For example, depending on the embodiment, the visual field or sensory field (e.g., Figures 4A-4CThe illustrated fields of view 404A, 404B, 404C, 404D, 404E) may together comprise a 360-degree field of view around the machine 800, or less than a 360-degree field of view around the machine 800. Furthermore, in some embodiments, one or more 3D signals, and therefore associated sensors, may comprise a 360-degree field of view or sensing field around the machine 800. For example, a rotating LiDAR sensor or a 360-degree surround camera may comprise a 360-degree field of view or sensing field, while a camera may comprise a 30-degree, 60-degree, 120-degree, and / or other field of view (e.g., up to 360 degrees, such as in a surround camera).

[0037] In cases where the sensor pipeline corresponds to a camera and / or other sensor type that does not directly compute depth, the 3D signal may still include predicted depth values represented in a rasterized image or other input representation. For example, the individual DNNs used in the sensor data pipeline may be trained to predict depth using a single camera image, or may be trained to predict depth from images with overlapping fields of view (e.g., such as may include images such as Figure 4A In some embodiments, the position prior image 114 may be generated as an additional input to the fused DNN 120, which indicates to the fused DNN 120 the predicted depth of the object based on the predicted depth from the 3D signal - or a probability distribution function 252 (represented as a 2D Gaussian representation in an embodiment) indicating the distribution of potential depths. For example, as shown in FIG. Figure 2C As shown, one or more sensors of the ego machine 800 can detect objects 240A, 240B, 240C, 240D, etc., as shown by the circles in the position prior image 114A. In such an example, where object position predictions are based on sensor data generated using a monocular camera or other sensor type where the accuracy of position predictions may be less than ideal, a distribution of potential positions can be fed to the fusion DNN 120 to help the fusion DNN 120 generate more accurate predictions in the fusion output 122. Thus, an ellipse or probability distribution function (PDF) 252 corresponding to each object 240 represents a potential location (e.g., with a corresponding confidence value) that can indicate where the object 240 may be located. For different sensor modalities, the corresponding ellipse or PDF 252 can have different shapes. For example, because the camera prediction can be along a ray (e.g., Figure 4A For example, the ellipse or PDF may be longer and narrower in shape (e.g., representing a bimodal distribution along the ray direction), while the RADAR sensor may have a shorter but wider representation to account for the corresponding inaccuracies or the RADAR sensor, etc. Figure 2DDifferent ellipses or PDFs 252 corresponding to RADAR sensor predictions within the RADAR sensor's field of view 250A may be shown. Furthermore, the shape of the ellipses or PDFs may vary depending on where the detection is located within the field of view of the corresponding sensor. For example, referring to Figure 2E As shown, the ellipse or PDF near the edge of field of view 250B or 250C or farther from the sensor may be different than at the center of field of view 250 or closer to the sensor - for example, at the edge, the prediction may not be as accurate as in the center, so the shape of the ellipse or PDF 252 may be larger to indicate more variability. Figure 2D A similar representation is shown in FIG for field of view 250A.

[0038] To generate a position prior image 114—e.g. Figure 2C The position prior image 114A or Figure 2E The position prior image 114B - a pixel or point corresponding to a detected object predicted using the sensor pipeline (e.g., corresponding to a quantized 3D position) can be used to retrieve a corresponding learned or determined ellipse or PDF corresponding to that point or pixel within the field of view of the corresponding sensor. For example, Figure 2D A subset of ellipses or PDFs may be represented for a particular sensor's field of view 250A. While the illustration includes multiple ellipses or PDFs 252A, 252B, 252C, etc., this is not intended to be limiting. For example, in some embodiments, each pixel or point representing the field of view may have a corresponding ellipse or PDF 252. In other embodiments, any number of pixels or points may have a corresponding ellipse or PDF 252. These ellipses or PDFs 252 may be learned or determined based on historical predictions from a particular sensor and the corresponding inaccuracies of these predictions. Thus, when a particular sensor detects an object at a particular point or pixel, the corresponding known ellipse or PDF 252 may be retrieved and inserted into the position prior image 114.

[0039] In some embodiments, as described herein, instead of or in addition to including one or more channels in the input signals of each pipeline, one or more channels can be used to generate separate input representations—e.g., velocity image(s) 116, instance / appearance image(s) 118, etc. For example, velocity information generated using one or more sensor pipelines can be used to generate one or more velocity images 116 and / or one or more instance / appearance images 118—e.g., temporal instance / appearance images and / or spatial instance / appearance images. Velocity image(s) 116 can include one or more channels corresponding to velocity in the x-direction, y-direction, etc. In some embodiments, as described herein, velocity image(s) 116 can be generated based on the output of one or more sensor pipelines. For example, one or more 3D signals from the sensor pipelines can include velocity information, and instead of or in addition to encoding this information in a rasterized image used as input to the fused DNN 120, the velocity information can be used to generate one or more separate inputs corresponding to velocity image(s) 116. Thus, velocity information may help the fusion DNN 120 determine whether objects or features in close proximity are the same object (e.g., similar or same velocity) or different objects (e.g., different velocity), and thus how to represent the objects in the fusion output 122—e.g., as a single object or as two or more objects.

[0040] As another example, instance / appearance image(s) 118 corresponding to instance and / or appearance spatial associations may be used. For example, one or more sensor pipelines may generate outputs indicating descriptors of features or objects. In such examples, one or more machine learning models in one or more sensor pipelines may be trained to generate an N-dimensional vector for each bounding shape / 2D cuboid / 3D cuboid of an object—for example, in embodiments, where N may be equal to 3. As a result, an N-dimensional vector may be generated for one or more detected objects or features. Thus, one or more vectors generated using one or more sensor pipelines may be compared to each other to determine how many object instances are present in a given frame. If a first vector and a second vector are sufficiently similar (e.g., within a threshold similarity), they may be determined to correspond to the same object, and this information may be represented in instance / appearance image 118. Similarly, instance and / or appearance images may be generated for temporal associations between objects or features. For example, a recurrent neural network may be used to take input vectors or instance / appearance information across frames and generate instance / appearance image 118 representing object instances that change over time. In either example, the instance / appearance image(s) 118 may be used as additional input(s) to the fusion DNN 120 in computing the fusion output 122. For example, the instance / appearance information may help the fusion DNN 120 determine whether objects or features in close proximity are the same object or different objects, and thus how to represent the objects in the fusion output 122—e.g., as a single object or as two or more objects.

[0041] At each iteration of the fused DNN 120, the 3D signals 104, 108, 112, etc., the position prior image(s) 114, the velocity image(s) 116, and / or the instance / appearance image(s) 118 may be provided as inputs to the fused DNN 120. The fused DNN 120 may process the inputs to generate a fused output 122, an example of which is shown in FIG. Figures 3A-3B The fused DNN 120 and / or one or more of the DNNs or machine learning models used to generate the 3D signal may include, for example, but not limited to, any type of machine learning model, such as one or more machine learning models using linear regression, logistic regression, decision trees, support vector machines (SVMs), natural Bayes, k-nearest neighbors (Knn), K-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short term memory (LSTM), Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid machines, etc.), and / or other types of machine learning models.

[0042] In embodiments where the fusion DNN 120 and / or one or more DNNs or machine learning models for generating (one or more) 3D signals include a convolutional neural network (CNN), one or more layers may include an input layer. The input layer may store values associated with various inputs. For example, when the input is (one or more) images (e.g., rasterized images corresponding to various channels from the 3D signal), the input layer may store values representing the raw pixel values of the (one or more) images as a volume (e.g., width W, height H, and color channels (e.g., RGB), such as 32x32x3), and / or a batch size B (e.g., when batch processing is used).

[0043] One or more layers can include a convolutional layer. A convolutional layer can compute the output of neurons connected to local regions in an input layer (e.g., an input layer), with each neuron computing the dot product between their weights and the small region they are connected to in the input volume. The result of the convolutional layer can be another volume where one dimension is based on the number of filters applied (e.g., width, height, and number of filters, such as 32 x 32 x 12 if 12 is the number of filters).

[0044] One or more layers may include a rectified linear unit (ReLU) layer. The ReLU layer may apply an element-by-element activation function, such as max(0, x), such as zero thresholding. The resulting volume of the ReLU layer may be the same as the volume of the input to the ReLU layer.

[0045] One or more layers may include a pooling layer. The pooling layer may perform a downsampling operation along the spatial dimensions (e.g., height and width), which may result in a smaller volume than the input to the pooling layer (e.g., 16 x 16 x 12 from a 32 x 32 x 12 input volume). In some examples, the fused DNN 120 and / or the other DNN(s) may not include any pooling layers. In such examples, strided convolutional layers may be used in place of pooling layers. In some examples, the fused DNN(s) 120 and / or the other DNN(s) (e.g., from Figure 6B The feature extractor layers (F1, F2, etc.) may include alternating convolutional layers and pooling layers, or may not include any pooling layers at all.

[0046] One or more layers may include fully connected layers. Each neuron in the (one or more) fully connected layers may be connected to each neuron in the previous volume. The fully connected layers may calculate class scores, and the resulting volume may be 1 x 1 x the number of classes. In some examples, the fused DNN 120 and / or the other (one or more) DNNs as a whole may not use fully connected layers in an effort to increase processing time and reduce computational resource requirements. In such examples, without using fully connected layers, the fused DNN 120 and / or the other (one or more) DNNs may be referred to as fully convolutional networks.

[0047] In some examples, one or more layers may include (one or more) deconvolution layers. However, the use of the term deconvolution may be misleading and is not intended to be limiting. For example, (one or more) deconvolution layers may alternatively be referred to as transposed convolution layers or fractionally strided convolution layers. (One or more) deconvolution layers may be used to perform upsampling on the output of a previous layer. For example, (one or more) deconvolution layers may be used to upsample to a spatial resolution equal to the spatial resolution of the input to the fusion DNN 120 and / or other (one or more) DNNs (e.g., the spatial resolution of a 3D signal), or to upsample to the input spatial resolution of the next layer.

[0048] Although input layers, convolutional layers, pooling layers, ReLU layers, deconvolutional layers, and fully connected layers are discussed herein with respect to DNN 120, this is not intended to be limiting. For example, additional or alternative layers such as normalization layers, SoftMax layers, and / or other layer types may be used.

[0049] Different orders and numbers of layers of the fused DNN 120 may be used depending on the embodiment. Furthermore, some layers may include parameters (e.g., weights and / or biases), while other layers may not include parameters, such as, for example, ReLU layers and pooling layers. In some examples, the parameters may be learned by the fused DNN 120 during training. Furthermore, some layers may include additional hyperparameters (e.g., learning rate, stride, epoch, kernel size, number of filters, pooling type for pooling layers, etc.), such as (one or more) convolutional layers, (one or more) deconvolutional layers, and (one or more) pooling layers, while other layers may not include them, such as (one or more) ReLU layers. Various activation functions may be used, including, but not limited to, ReLU, leaky ReLU, sigmoid, hyperbolic tangent (tanh), exponential linear unit (ELU), etc. Parameters, hyperparameters, and / or activation functions are not limited and may vary depending on the embodiment.

[0050] In some embodiments, such as Figure 6BAs described, one or more DNNs for generating a 3D signal may include a backbone T, a feature extractor layer F, and / or a head H that may be used to generate the 3D signals 104, 108, 112, etc. In examples, and as described herein, in addition to or in lieu of using the 3D signals 104, 108, 112, etc. generated from various input sources, the fused DNN 120 may directly access the output of the feature extractor layer (e.g., Figure 1A F1-F n ) as input to generate the fused output 122. In such an example, the feature extractor layers of the individual DNNs may be directly connected to one or more layers of the fused DNN(s) 120, so that during training, backpropagation from the fused DNN 120 may be passed to the individual DNN(s) to train the individual DNN(s) and the fused DNN 120 together.

[0051] The fused DNN 120 may use the output of one or more layers of the (one or more) individual DNNs, the 3D signals 104, 108, 112, etc., the (one or more) position prior images 114, the (one or more) velocity images 116, and / or the instance appearance image 118 to generate a fused output 122. In embodiments, the fused output 122 may be similar to the 3D signal—for example, the fused output 122 may include a representation or may be used to generate a rasterized image (e.g., similar to Figures 3A-3B For example, the fused output may represent a top-down bird's-eye view image, a projected image such as a range image, a side view image, etc., encoding an arbitrary number of output channels (e.g., similar to Figures 1A-1B In embodiments where the ego machine 800 is located at the center of the representation, the fused output rasterized image computed using the fused DNN may be centered on the ego. The fused output 122 may include a fused representation of one or more visual or sensory fields of one or more input sources. For example, Figure 2A The rasterized image of includes the field of view of the forward-facing camera and Figure 2B In the case where the rasterized image includes the field of view of the rear-facing camera, Figure 3A The fusion output 300A can be expressed as Figure 2A The rasterized image and Figure 2B The rasterized image of both. Figure 3B Another visualization of the fused output 300B may be included that includes the objects 206 (including the object 206D) and the lanes 204 (including the lanes 204C and 204D).

[0052] Figure 2A 、 2BThe illustrations of 3D and 3A are for example purposes only and are not intended to be limiting. For example, depending on the embodiment, the 3D signals may represent smaller and / or larger portions of the surrounding environment of the ego vehicle 800 and may represent a fusion of any number of input sources. Furthermore, one or more 3D signals may include a field of view or sensory field that overlaps with the field of view or sensory field of one or more other 3D signals, such that the fusion DNN 120 may learn to more accurately predict distances using, for example, stereo camera capabilities.

[0053] The fused output 122 can be used by the autonomous driving software stack ("driver stack") 124 to perform one or more operations of the vehicle 800 (and / or other ego machine types). For example, the driver stack 124 can include a world model manager that can be used to generate, update, and / or define a world model. The world model manager can use information generated and received by the perception component(s) of the driver stack 124. The perception component(s) can include obstacle sensors, path sensors, wait sensors, map sensors, and / or other perception component(s). For example, the world model can be defined at least in part based on the availability of obstacles, paths, and wait conditions, which can be sensed in real time or near real time by the obstacle sensors, path sensors, wait sensors, and / or map sensors. The world model manager can continuously update the world model based on newly generated and / or received input (e.g., data) from the obstacle sensors, path sensors, wait sensors, map sensors, and / or other components of the ego machine 800. For example, the world model manager and / or perception component may use the fused output 122 to perform one or more operations.

[0054] The world model may be used to help inform the planning component(s), control component(s), obstacle avoidance component(s), and / or actuation component(s) of the drive stack 124. An obstacle sensor may perform obstacle perception, which may be based on where the vehicle 800 is permitted to drive or is capable of driving, and how fast the vehicle 800 may travel without colliding with obstacles (e.g., objects such as structures, solids, vehicles, etc.) sensed by the vehicle 800 (and represented, for example, in the fused output 122).

[0055] Path perception can perform path perception, for example, by sensing a nominal path available in a particular situation. In some examples, path perception can further consider lane changes for path perception. A lane map can represent one or more paths available to vehicle 800 and can be as simple as a single path on a highway entrance ramp. In some examples, a lane map can include a path to a desired lane and / or can indicate available changes along a highway (or other road type), or can include nearby lanes, lane changes, forks, turns, interchanges, merges, and / or other information.

[0056] The waiting sensor may be responsible for determining constraints on vehicle 800 due to rules, conventions, and / or practical considerations. For example, rules, conventions, and / or practical considerations may relate to traffic lights, multi-way stops, yields, merges, toll booths, gates, police or other emergency personnel, road crews, stopped buses or other vehicles, one-way bridge arbitration, ferries, and the like. In some examples, the waiting sensor may be responsible for determining longitudinal constraints on vehicle 800 that require the vehicle to wait or slow down until a certain condition is true. In some examples, for example, the waiting condition is caused by a potential obstacle, such as traffic at an intersection, which may not be perceived directly by the obstacle sensor (e.g., by using sensor data from a sensor because the obstacle may be obscured from the sensor's field of view). As a result, the waiting sensor can provide situational awareness by addressing the danger of obstacles that are not always immediately perceived through rules and conventions that can be sensed and / or learned. Thus, the waiting sensor can be utilized to identify potential obstacles and implement one or more controls (e.g., slow down, stop, etc.) that may not be possible using the obstacle sensor alone.

[0057] A map sensor may include mechanisms for identifying behaviors and, in some examples, determining specific instances of what conventions apply in a particular location.

[0058] The planning component(s) may include a route planner, a lane planner, a behavior planner, and a behavior selector, among other components, features, and / or functionality. The route planner may use information from the map sensor, the map manager, and / or the positioning manager, among other information, to generate a planned path that may be comprised of GNSS waypoints (e.g., GPS waypoints). A waypoint may represent a specific distance into the future for the vehicle 800, such as a number of city blocks, a number of kilometers / miles, a number of meters / feet, etc., which may be used as a target for the lane planner.

[0059] The lane planner can use a lane map (e.g., a lane map from a path sensor, which can be generated at least in part using the fusion output 122), object poses in the lane map (e.g., according to the localization manager), and / or a target point and direction from the route planner to the future distance as input. The target point and direction can be mapped to the best matching drivable point and direction in the lane map (e.g., based on GNSS and / or compass direction). A graph search algorithm can then be performed on the lane map to find the shortest path to the target point from the current edge in the lane map.

[0060] The behavior planner can determine the feasibility of basic behaviors of the vehicle 800, such as staying in the lane or changing lanes left or right, so that the feasible behaviors can be matched with the most desired behavior output from the lane planner. For example, if the desired behavior is determined to be unsafe and / or unavailable, a default behavior can be selected instead (for example, when the desired behavior or lane change is unsafe, the default behavior may be staying in the lane).

[0061] The control component(s) may follow the trajectory or path (lateral and longitudinal) received from the behavior selector of the planning component to the extent possible and within the capabilities of the vehicle 800 .

[0062] The obstacle avoidance component(s) may help the autonomous vehicle 800 avoid collisions with objects (e.g., moving objects and stationary objects). In some examples, the obstacle avoidance component(s) may be used independently of components, features, and / or functions of the vehicle 800 that require compliance with traffic laws and courteous driving. In such examples, the obstacle avoidance component(s) may ignore traffic laws, road rules, and courteous driving norms to ensure that no collision occurs between the vehicle 800 and any object. Thus, the obstacle avoidance layer may be a layer independent of the road layer rules, and the obstacle avoidance layer may ensure that the vehicle 800 performs safe maneuvers only from an obstacle avoidance perspective. On the other hand, the road layer rules may ensure that the vehicle complies with traffic laws and conventions and adheres to legal and customary rights of way.

[0063] In any example, one or more layers, components, features, and / or functions of the drive stack 124 can use the fused output 122 to generate outputs for world model management, planning, control, actuation, collision or obstacle avoidance, etc. to help the self-machine 800 navigate the environment.

[0064] Now refer to Figure 5, each block of the method 500 described herein comprises a computing process that can be performed using any combination of hardware, firmware, and / or software. For example, various functions can be performed by a processor executing instructions stored in a memory. The method 500 can also be embodied as computer-usable instructions stored on a computer storage medium. The method 500 can be provided by a standalone application, a service, or a hosted service (either standalone or in combination with another hosted service), or a plug-in to another product, to name a few. In addition, as an example, for Figures 1A-1B The method 500 is described with reference to processes 100A and 100B. However, the method 500 may additionally or alternatively be performed by any system or process and any combination of systems within any process, including but not limited to those described herein.

[0065] Figure 5 is a flow chart illustrating a method 500 for computing a fused output using multiple input channels using a multi-sensor fusion network, according to some embodiments of the present disclosure. At block B 502, the method 500 includes receiving first data representing at least a first rasterized image, the first rasterized image generated at least in part based on first sensor data, the first sensor data generated using a first sensor of a first type. For example, a first sensor pipeline may be used to generate first 3D signals 104, 108, 112, etc.

[0066] At block B504, method 500 includes receiving second data representing at least a second rasterized image, the second rasterized image generated at least in part based on second sensor data, the second sensor data generated using a second sensor of a second type. For example, a second sensor pipeline may be used to generate the second 3D signals 104, 108, 112, etc.

[0067] At block B506, method 500 includes using the fused DNN and computing third data representing a fused rasterized image based at least in part on the first data and the second data. For example, fused DNN 120 may use the first 3D signal, the second 3D signal, and / or one or more other 3D signals to compute fused output 122. In some embodiments, as described herein, in addition to or in lieu of using the 3D signals as input, fused DNN 120 may use the output of one or more feature extractor layers of one or more separate DNNs corresponding to one or more sensor pipelines.

[0068] At block B508 , the method 500 includes using the autonomous machine to perform one or more operations based at least in part on the third data. For example, the driver stack 124 of the autonomous machine 800 may use the fused output 122 to perform one or more operations, such as world model management, planning, control, actuation, obstacle avoidance, etc.

[0069] refer to Figures 6A-6B , Figures 6A-6B Together, they represent a data flow diagram of a process 600 for training a multi-sensor fusion network and one or more associated source networks according to some embodiments of the present disclosure. Figure 6B , a separate sensor pipeline may include a separate or source DNN that includes, for example, one or more backbone layers 604 (e.g., backbone layer(s) 604A, 604B, and 604N), one or more feature extractor layers 606 (e.g., feature extractor layers 606A, 606B, and 606N), and / or one or more head or output layers 608 (e.g., output heads 608A, 608B, and 608N). A sensor pipeline may include sensors 602 (e.g., sensors 602A, 602B, and 602N) that generate sensor data—e.g., as described herein for Figures 1A-1B Any sensor data described - which can be processed using the source DNN to generate a 3D signal (e.g., 3D signals 104A, 104B, and 104N). For example, the sensors may include cameras, LiDAR sensors, RADAR sensors, ultrasonic sensors, and / or other sensor types that can be used to generate sensor data for the respective DNNs. Using one or more loss functions for calculating a loss 612 corresponding to the source DNN, the 3D signal output by the source DNN can be compared to ground truth data 614. The ground truth data 614 can be generated using map data (e.g., from an HD map or other map type, such as those used for positioning), which map data can indicate the location of static features or objects, such as lane lines, waiting conditions, signs, fixed objects, and / or the like. In some embodiments, the ground truth 614 can be generated using a 2D or 3D ground truth generation pipeline, such as that described in U.S. non-provisional application No. 17 / 187,350, filed on February 26, 2021, which is incorporated herein by reference in its entirety. Ground truth data 614 may include rasterized images, including information about Figures 1A-1B 104, 108, 112, etc. in the source DNN. For example, the output 3D signal 104 of the source DNN may correspond to a rasterized image, or may be used to generate a rasterized image, and the rasterized image may be compared with the ground truth rasterized image to calculate losses 622 (e.g., losses 622A, 622B, and 622N). These losses may be used to update the parameters (e.g., weights and biases) of the source DNN using, for example, backpropagation to help train the DNN(s) until they converge to an acceptable level of accuracy or precision.

[0070] In addition to or in lieu of calculating the loss 622 using the ground truth data 614, a photometric consistency loss can be calculated at 610 to compare the output 3D signals 104 from two or more source DNNs of the sensor 602 having at least partially overlapping fields of view. For example, similar to generating a disparity map for a stereo camera, a photometric consistency loss 624 (e.g., a loss 624A between the 3D signal 104A and the 3D signal 104B and / or between the 3D signal 104B and the 3D signal 104A, a loss 624B between the 3D signal 104N and the 3D signal 104K and / or between the 3D signal 104K and the 3D signal 104N, etc.) can be calculated to compare the outputs of the two or more source DNNs to ensure that at least partially overlapping fields of view (e.g., Figure 4A 406). Thus, with respect to 3D signal 104A and 3D signal 104B, coordinate converter 618 may convert 3D signal 104A to the coordinate space of 3D signal 104B to generate 3D signal 616A, and then calculate consistency loss 624A by comparing the converted 3D signal 616A with 3D signal 104B to determine consistency or lack thereof between the signals. Thus, in cases where the overlapping region is inconsistent—for example, where an object in converted 3D signal 616A is different from the same object in 3D signal 104B (e.g., in terms of position, depth, orientation, category, etc.), loss 624A may be higher and the source DNN corresponding to 3D signal 104A and / or 3D signal 104B may be penalized (e.g., parameters may be updated). This process may be similarly performed by converting 3D signal 104B into the coordinate space of 3D signal 104A using coordinate converter 618 to generate 3D signal 616B, converting 3D signal 104N into the coordinate space of 3D signal 104K using coordinate converter 618 to generate 3D signal 616C, converting 3D signal 104K into the coordinate space of 3D signal 104N using coordinate converter 618 to generate 3D signal 616D, etc. Thus, by calculating loss 622 and / or loss 624, the source DNNs may be trained until they reach an acceptable level of accuracy or precision.

[0071] One benefit of training the source DNN(s) separately from the fused DNN 120 is that the sensor data or simulated sensor data used in training need not rely on realistic renderings of the input images. For example, because the 3D signals 104, 108, 112, etc. may represent rasterized images generated using various channels, rasterized images for training may be generated without the need for realistic sensor data—e.g., because the input to the fused DNN 120 is a rasterized image.

[0072] Similarly, and with respect to Figure 6A and6B , the fused output 122 of the fused DNN 120 can be compared to the ground truth data 614 corresponding to the fused output to calculate a loss 630. The ground truth data 614 of the fused output 122 can be generated using similar techniques or data as the ground truth data 614 of the source DNN. Therefore, the loss 630 can be used to update the parameters of the fused DNN 120 until the fused DNN 120 converges to an acceptable level of accuracy or precision. Although other input types - such as the position prior channel 114, (one or more) velocity images 116 and / or (one or more) instance / appearance images 118 are not shown in FIG. Figures 6A-6B , but this is not intended to be limiting, and in some embodiments, one or more of these (one or more) input channels may also be provided as input to the fused DNN 120 at each training iteration.

[0073] For example, and about Figure 6C , the fused output 646 may include the object 648, and the ground truth data 614 may indicate the actual location of the object as the ground truth object 650. In such an example, the calculated loss 630 may represent this difference, and the parameters of the fused DNN 120 may be updated. In examples where the 3D signal from the source DNN corresponds to a rasterized image (e.g., including less or an equal amount of the surrounding environment of the ego machine 800), the loss 622 may be similarly calculated using the corresponding rasterized image of the 3D signal.

[0074] In some embodiments, as described herein, the inputs to the fused DNN 120 may correspond to the feature outputs 102—e.g., the outputs of the feature extractor layer 606 of the source DNN—such that the layers of the source DNN may correspond to separate input trunks or layers for the fused DNN 120. In such examples, connections may be connected between nodes of the feature extractor layer 606 of the source DNN and nodes of a layer (e.g., an input layer) of the fused DNN 120. In this way, the fused DNN 120 and the source DNNs may be trained simultaneously such that the loss computed for the fused DNN 120 may be back-propagated to the source DNNs.

[0075] Referring now to FIG6 , each block of the method 700 described herein comprises a computing process 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 method 700 may also be embodied as computer-usable instructions stored on a computer storage medium. The method 700 may be provided by a standalone application, a service, or a hosted service (either standalone or in combination with another hosted service), or a plug-in to another product, to name a few. Furthermore, as an example, the method 700 may be provided with respect to Figures 6A-6BHowever, the method 700 may additionally or alternatively be performed by any system or process and any combination of systems within any process, including but not limited to those described herein.

[0076] Figure 7 7 is a flow chart illustrating a method 700 for training a multi-sensor fusion network to compute a fusion output using multiple input channels, according to some embodiments of the present disclosure. At block B702, the method 700 includes generating a plurality of trained DNNs by computing one or more losses on the outputs of individual DNNs and computing one or more consistency losses on the outputs of two or more individual DNNs. For example, a source DNN may be trained using ground truth data 614 to generate a loss 622, and a photometric consistency loss 624 may be used to train the source DNN.

[0077] At block B704, method 700 includes generating a trained fused DNN by computing one or more losses with respect to the fused outputs, the fused outputs being computed using the outputs of the plurality of trained DNNs. For example, once trained, the source DNNs may compute outputs, and these outputs may be used as inputs to the fused DNN 120 to train the fused DNN 120. For example, the fused DNN 120 may compute (one or more) fused outputs 122 and may compare (one or more) fused outputs 122 with the ground truth data 614 to compute (one or more) losses 630. In some examples, the fused DNN 120 may additionally or alternatively be trained using simulated or manufactured data (e.g., rasterized images) without using the actual outputs of the source DNNs.

[0078] At block B706, method 700 includes deploying the plurality of trained DNNs and the fused DNN in the ego machine. For example, once trained, the fused DNN 120 and the source DNN may be deployed in the ego machine 800.

[0079] Example autonomous vehicle

[0080] Figure 8A8 is an illustration of an example autonomous vehicle 800 according to some embodiments of the present disclosure. Autonomous vehicle 800 (alternatively referred to herein as "vehicle 800") 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 motorized bicycle, a motorcycle, a fire truck, a police vehicle, an ambulance, a boat, a construction vehicle, an underwater vessel, a drone, a vehicle coupled to a trailer, and / or another type of vehicle (e.g., a vehicle that is unmanned and / or accommodates one or more passengers). Autonomous vehicles are generally described according to the levels of automation 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, issued on June 15, 2018, Standard No. J3016-201609, issued on September 30, 2016, and previous and future versions of such standards). The vehicle 800 may be capable of implementing functionality consistent with one or more of autonomous driving levels 3 to 5. For example, depending on the embodiment, the vehicle 800 may be capable of conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0081] Vehicle 800 may include components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other vehicle components. Vehicle 800 may include a propulsion system 850, such as an internal combustion engine, a hybrid power plant, an all-electric engine, and / or another type of propulsion system. Propulsion system 850 may be connected to a drivetrain of vehicle 800, which may include a transmission, to achieve propulsion of vehicle 800. Propulsion system 850 may be controlled in response to receiving a signal from throttle / accelerator 852.

[0082] A steering system 854, which may include a steering wheel, may be used to steer the vehicle 800 (e.g., along a desired path or route) when the propulsion system 850 is operating (e.g., when the vehicle is in motion). The steering system 854 may receive signals from a steering actuator 856. For fully automated (Level 5) functionality, a steering wheel may be optional.

[0083] Brake sensor system 846 may be used to operate vehicle brakes in response to receiving signals from brake actuator 848 and / or brake sensors.

[0084] May include one or more system on chip (SoC) 804 ( Figure 8C) and / or one or more GPUs can provide signals (e.g., representing commands) to one or more components and / or systems of the vehicle 800. For example, the one or more controllers can send signals to operate the vehicle brakes via one or more brake actuators 848, to operate the steering system 854 via one or more steering actuators 856, and to operate the propulsion system 850 via one or more throttles / accelerators 852. The one or more controllers 836 can include one or more onboard (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 assist a human driver in driving the vehicle 800. The one or more controllers 836 can include a first controller 836 for autonomous driving functions, a second controller 836 for functional safety functions, a third controller 836 for artificial intelligence functions (e.g., computer vision), a fourth controller 836 for infotainment functions, a fifth controller 836 for redundancy in emergency situations, and / or other controllers. In some examples, a single controller 836 may handle two or more of the above functions, two or more controllers 836 may handle a single function, and / or any combination thereof.

[0085] The one or more controllers 836 may provide signals for controlling one or more components and / or systems of the vehicle 800 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, a global navigation satellite system sensor 858 (e.g., a global positioning system sensor), a RADAR sensor 860, an ultrasonic sensor 862, a LIDAR sensor 864, an inertial measurement unit (IMU) sensor 866 (e.g., an accelerometer, a gyroscope, a magnetic compass, a magnetometer, etc.), a microphone 896, a stereo camera 868, a wide-angle camera 870 (e.g., a fisheye camera), an infrared camera 872, a surround camera 874 (e.g., a 360-degree camera), a long-range and / or mid-range camera 898, a speed sensor 844 (e.g., for measuring the velocity of the vehicle 800), a vibration sensor 842, a steering sensor 840, a brake sensor (e.g., as part of a brake sensor system 846), and / or other sensor types.

[0086] One or more of the controllers 836 may receive input (e.g., represented by input data) from the instrument cluster 832 of the vehicle 800 and provide output (e.g., represented by output data, display data, etc.) via a human machine interface (HMI) display 834, an audible annunciator, a speaker, and / or via other components of the vehicle 800. These outputs may include information such as vehicle speed, velocity, time, map data (e.g., Figure 8CThe HMI display 834 may include information such as the HD map 822 of the vehicle 800, location data (e.g., the location of the vehicle 800 on the map), directions, the locations of other vehicles (e.g., an occupancy grid), information about objects and object states as sensed by the controller 836, etc. For example, the HMI display 834 may display information about the presence of one or more objects (e.g., street signs, warning signs, traffic light changes, etc.) and / or information about driving maneuvers that the vehicle has made, is making, or will make (e.g., changing lanes now, leaving 34B in two miles, etc.).

[0087] The vehicle 800 also includes a network interface 824 that can communicate over one or more networks using one or more wireless antennas 826 and / or a modem. For example, the network interface 824 can be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The one or more wireless antennas 826 can also enable communication between objects in the environment (e.g., vehicles, mobile devices, etc.) using one or more local area networks such as Bluetooth, Bluetooth LE, Z-wave, ZigBee, etc. and / or one or more low power wide area networks (LPWANs) such as LoRaWAN, SigFox, etc.

[0088] Figure 8B For use according to some embodiments of the present disclosure Figure 8A An example of camera positions and fields of view for autonomous vehicle 800 is shown. The cameras and respective fields of view are an example embodiment and are not intended to be limiting. For example, additional and / or alternative cameras may be included and / or located at different locations on vehicle 800.

[0089] The camera type used for the camera may include, but is not limited to, a digital camera that may be suitable for use with components and / or systems of the vehicle 800. The camera may operate at Automotive Safety Integrity Level (ASIL) B and / or at another ASIL. The camera type may have any image capture rate, such as 60 frames per second (fps), 120fps, 240fps, and the like, 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 examples, the color filter array may include a red-white-white-white (RCCC) color filter array, a red-white-white-blue (RCCB) color filter array, a red-blue-green-white (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, a clear pixel camera such as a camera with an RCCC, RCCB, and / or RBGC color filter array may be used in an effort to improve light sensitivity.

[0090] In some examples, one or more of the cameras can 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 monocular camera can be installed to provide functions including lane departure warning, traffic sign assistance, and intelligent headlight control. One or more of the cameras (e.g., all of the cameras) can simultaneously record and provide image data (e.g., video).

[0091] One or more of the cameras can be mounted in a mounting assembly, such as a custom-designed (3-D printed) assembly, to cut off stray light and reflections from within the car (e.g., reflections from the dashboard reflected in the windshield mirror) that could interfere with the camera's ability to capture image data. With respect to the wing mirror mounting assembly, the wing mirror assembly can be custom 3-D printed so that the camera mounting plate matches the shape of the wing mirror. In some examples, one or more cameras can be integrated into the wing mirror. For side-view cameras, one or more cameras can also be integrated into the four pillars at each corner of the cab.

[0092] A camera (e.g., a front-facing camera) having a field of view that includes a portion of the environment in front of the vehicle 800 can be used for surround vision to help identify the forward path and obstacles, as well as assist in providing information critical to generating an occupancy grid and / or determining a preferred vehicle path with the help of one or more controllers 836 and / or control SoCs. The front-facing camera can be used to perform many of the same ADAS functions as LIDAR, including emergency braking, pedestrian detection, and collision avoidance. The front-facing camera can also be used for ADAS functions and systems, including lane departure warning (LDW), autonomous cruise control (ACC), and / or other functions such as traffic sign recognition.

[0093] A variety of cameras can be used in the front-facing configuration, including, for example, a monocular camera platform including a CMOS (complementary metal oxide semiconductor) color imager. Another example can be a wide-angle camera 870, which can be used to sense objects entering the field of view from the periphery (e.g., pedestrians, intersection traffic, or bicycles). Although Figure 8B The figure shows only one wide-angle camera, but there can be any number of wide-angle cameras 870 on the vehicle 800. In addition, long-range cameras 898 (e.g., a long-view stereo camera pair) can be used for depth-based object detection, especially for objects for which neural networks have not yet been trained. Long-range cameras 898 can also be used for object detection and classification and basic object tracking.

[0094] One or more stereo cameras 868 may also be included in the front configuration. The stereo camera 868 may include an integrated control unit including an expandable processing unit that may provide a multi-core microprocessor and programmable logic (FPGA) with an integrated CAN or Ethernet interface on a single chip. Such a unit may be used to generate a 3-D map of the vehicle environment, including distance estimates for all points in the image. An alternative stereo camera 868 may include a compact stereo vision sensor that may include two camera lenses (one on the left and one on the right) and an image processing chip that may measure the distance from the vehicle to the target object and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 868 may be used in addition to or alternatively to those described herein.

[0095] Cameras with a field of view that includes portions of the environment to the sides of the vehicle 800 (e.g., side-view cameras) can be used for surround viewing, providing information used to create and update occupancy grids and generate side impact collision warnings. For example, surround cameras 874 (e.g., Figure 8B Four surround cameras 874 (shown in FIG) can be placed on the vehicle 800. The surround cameras 874 can include a wide-angle camera 870, a fisheye camera, a 360-degree camera, and / or the like. For example, the four fisheye cameras can be placed on the front, rear, and sides of the vehicle. In an alternative arrangement, the vehicle can use three surround cameras 874 (e.g., left, right, and rear), and can utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround-view camera.

[0096] A camera having a field of view that includes a portion of the environment behind the vehicle 800 (e.g., a rearview camera) can be used to assist with parking, surround view, rear collision warning, and creating and updating occupancy grids. A variety of cameras can be used, including but not limited to cameras that are also suitable as front-facing cameras as described herein (e.g., long-range and / or mid-range cameras 898, stereo cameras 868, infrared cameras 872, etc.).

[0097] Figure 8C For use according to some embodiments of the present disclosure Figure 8A800. It will be understood that this arrangement and other arrangements described herein are set forth merely as examples. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, functional groupings, etc.) may be used in addition to or in place of those shown, and some elements may be omitted entirely. Further, many of the elements described herein are functional entities that may be implemented as discrete or distributed components or in conjunction with other components, and in any appropriate combination and location. The various functions described herein as being performed by entities may be implemented by hardware, firmware, and / or software. For example, the various functions may be implemented by a processor executing instructions stored in memory.

[0098] Figure 8C Each of the components, features, and systems of vehicle 800 is illustrated as being connected via bus 802. Bus 802 may include a controller area network (CAN) data interface (alternatively, referred to herein as a "CAN bus"). CAN may be a network internal to vehicle 800 that assists in controlling various features and functions of vehicle 800, such as actuation of brakes, acceleration, braking, steering, windshield wipers, and the like. The CAN bus may be configured to have tens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus may be read to find steering wheel angle, ground speed, engine speed per minute (RPM), button position, and / or other vehicle status indicators. The CAN bus may be ASIL B compliant.

[0099] Although bus 802 is described here as a CAN bus, this is not intended to be limiting. For example, FlexRay and / or Ethernet may be used in addition to or in lieu of a CAN bus. Furthermore, although bus 802 is represented by a single line, this is not intended to be limiting. For example, there may be any number of buses 802, which may include one or more CAN buses, one or more FlexRay buses, one or more Ethernet buses, and / or one or more other types of buses using different protocols. In some examples, two or more buses 802 may be used to perform different functions and / or may be used for redundancy. For example, a first bus 802 may be used for collision avoidance functionality, and a second bus 802 may be used for drive control. In any example, each bus 802 may communicate with any component of vehicle 800, and two or more buses 802 may communicate with the same component. In some examples, each SoC 804, each controller 836, and / or each computer within the vehicle may have access to the same input data (e.g., input from sensors of vehicle 800) and may be connected to a common bus such as a CAN bus.

[0100] The vehicle 800 may include one or more controllers 836, such as those described herein. Figure 8A Controller 836 may be used for a variety of functions. Controller 836 may be coupled to any of the other various components and systems of vehicle 800 and may be used for control of vehicle 800, artificial intelligence of vehicle 800, infotainment for vehicle 800, and / or the like.

[0101] The vehicle 800 may include one or more system-on-chips (SoCs) 804. The SoC 804 may include a CPU 806, a GPU 808, a processor 810, a cache 812, an accelerator 814, a data store 816, and / or other components and features not shown. The SoC 804 may be used to control the vehicle 800 in a variety of platforms and systems. For example, the one or more SoCs 804 may be combined with an HD map 822 in a system (e.g., a system of the vehicle 800), which may be downloaded from one or more servers (e.g., a server) via a network interface 824. Figure 8D one or more servers 878) to obtain map refreshes and / or updates.

[0102] The CPU 806 may include a CPU cluster or CPU complex (alternatively, referred to herein as a "CCPLEX"). The CPU 806 may include multiple cores and / or L2 caches. For example, in some embodiments, the CPU 806 may include eight cores in a coherent multiprocessor configuration. In some embodiments, the CPU 806 may include four dual-core clusters, each with a dedicated L2 cache (e.g., a 2MB L2 cache). The CPU 806 (e.g., CCPLEX) may be configured to support simultaneous cluster operations such that any combination of CPU 806 clusters can be active at any given time.

[0103] The CPU 806 may implement power management capabilities including one or more of the following features: each hardware block 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 due to the 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 806 may further implement an enhanced algorithm for managing power states, in which allowed power states and expected wakeup times are specified, and hardware / microcode determines the optimal power state to enter for the core, cluster, and CCPLEX. The processing core may support a simplified power state entry sequence in software, with this work being offloaded to the microcode.

[0104] The GPU 808 may include an integrated GPU (alternatively referred to herein as an "iGPU"). The GPU 808 may be programmable and efficient for parallel workloads. In some examples, the GPU 808 may use an enhanced tensor instruction set. The GPU 808 may include one or more streaming microprocessors, each of which may include an L1 cache (e.g., an L1 cache with at least 96KB of storage capacity), and two or more of these streaming microprocessors may share an L2 cache (e.g., an L2 cache with 512KB of storage capacity). In some embodiments, the GPU 808 may include at least eight streaming microprocessors. The GPU 808 may use a computing application programming interface (API). In addition, the GPU 808 may use one or more parallel computing platforms and / or programming models (e.g., NVIDIA's CUDA).

[0105] In the case of automotive and embedded use, GPU 808 can be power optimized to achieve optimal performance. For example, GPU 808 can be manufactured on fin field effect transistors (FinFETs). However, this is not intended to be limiting, and GPU 808 can be manufactured using other semiconductor manufacturing processes. Each streaming microprocessor can merge several mixed precision processing cores divided into multiple blocks. For example and without limitation, 64 PF32 cores and 32 PF64 cores can be divided into four processing blocks. In such an example, each processing block can be allocated 16 FP32 cores, 8 FP64 cores, 16 INT32 cores, two mixed precision NVIDIA tensor cores for deep learning matrix arithmetic, L0 instruction cache, warp scheduler, dispatch unit and / or 64KB register file. In addition, the streaming microprocessor may include independent parallel integer and floating point data paths to provide efficient execution of workloads using a mix of computation and addressing calculations. The streaming microprocessor may include independent thread scheduling capabilities to allow for finer-grained synchronization and collaboration between parallel threads. Streaming microprocessors may include a combined L1 data cache and shared memory unit to increase performance while simplifying programming.

[0106] The GPU 808 can include high bandwidth memory (HBM) and / or a 16GB HBM2 memory subsystem that provides a peak memory bandwidth of approximately 900 GB / s in some examples. In some examples, synchronous graphics random access memory (SGRAM), such as fifth generation graphics double data rate synchronous random access memory (GDDR5), can be used in addition to or in lieu of HBM memory.

[0107] The GPU 808 may include unified memory technology that includes access counters to allow memory pages to be more accurately migrated to the processor that accesses them most frequently, thereby improving the efficiency of memory ranges shared between processors. In some examples, address translation services (ATS) support may be used to allow the GPU 808 to directly access the CPU 806 page tables. In such an example, when the GPU 808 memory management unit (MMU) experiences a miss, an address translation request may be transmitted to the CPU 806. In response, the CPU 806 may look up the virtual-to-physical mapping for the address in its page table and transmit the translation back to the GPU 808. In this way, unified memory technology may allow a single unified virtual address space to be used for memory of both the CPU 806 and the GPU 808, thereby simplifying GPU 808 programming and porting applications to the GPU 808.

[0108] In addition, GPU 808 can include access counters that can track how often GPU 808 accesses the memory of other processors. Access counters can help ensure that memory pages are moved to the physical memory of the processor that accesses them most frequently.

[0109] SoC 804 may include any number of caches 812, including those described herein. For example, cache 812 may include an L3 cache available to both CPU 806 and GPU 808 (e.g., connected to both CPU 806 and GPU 808). Cache 812 may include a write-back cache that can track the state of lines, for example, using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). Depending on the embodiment, the L3 cache may include 4MB or more, although smaller cache sizes may also be used.

[0110] The SoC 804 may include an arithmetic logic unit (ALU) that may be utilized in performing any of a variety of tasks or operations associated with the vehicle 800, such as processing a DNN. Furthermore, the SoC 804 may include a floating point unit (FPU) (or other math coprocessor or digital coprocessor type) for performing mathematical operations within the system. For example, the SoC 104 may include one or more FPUs integrated as execution units within the CPU 806 and / or GPU 808.

[0111] SoC 804 may include one or more accelerators 814 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, SoC 804 may include a hardware accelerator cluster, which may include optimized hardware accelerators and / or large on-chip memory. The large on-chip memory (e.g., 4MB SRAM) may enable the hardware accelerator cluster to accelerate neural networks and other calculations. The hardware accelerator cluster may be used to supplement the GPU 808 and offload some tasks of the GPU 808 (e.g., freeing up more cycles of the GPU 808 for performing other tasks). As an example, the accelerator 814 may be used for targeted workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are sufficiently stable to be easily controlled for acceleration. When used herein, the term "CNN" may include all types of CNNs, including region-based or regional convolutional neural networks (RCNNs) and fast RCNNs (e.g., for object detection).

[0112] The accelerator 814 (e.g., a hardware accelerator 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 reasoning. The TPU may be an accelerator configured to perform image processing functions (e.g., for CNN, RCNN, etc.) and optimized for performing image processing functions. The DLA may be further optimized for a specific set of neural network types and floating-point operations and reasoning. The design of the DLA may provide higher performance per millimeter than a general-purpose GPU and far exceed the performance of the CPU. The TPU may perform several functions, including a single-instance convolution function, support for INT8, INT16, and FP16 data types for both features and weights, and post-processor functions.

[0113] DLA can quickly and efficiently execute neural networks, particularly CNNs, on processed or unprocessed data for any of a wide variety of functions, such as, but not limited to: CNNs for object recognition and detection using data from camera sensors; CNNs for distance estimation using data from camera sensors; CNNs for emergency vehicle detection and recognition and detection using data from microphones; CNNs for facial recognition and vehicle owner identification using data from camera sensors; and / or CNNs for safety and / or security-related events.

[0114] The DLA can perform any function of the GPU 808, and by using an inference accelerator, for example, the designer can target any function to either the DLA or the GPU 808. For example, the designer can focus the processing of CNNs and floating-point operations on the DLA and leave other functions to the GPU 808 and / or other accelerators 814.

[0115] The accelerator 814 (e.g., a hardware accelerator cluster) may include a programmable vision accelerator (PVA), which may be alternatively 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, for example and without limitation, any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors.

[0116] The RISC core can interact with an image sensor (e.g., an image sensor of any camera described herein), an image signal processor, and / or the like. Each of these RISC cores can include any amount of memory. Depending on the embodiment, the RISC core can use any of a number of protocols. In some examples, the RISC core can execute a real-time operating system (RTOS). The RISC core can be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC core can include an instruction cache and / or tightly coupled RAM.

[0117] The DMA can enable components of the PVA to access system memory independently of the CPU 806. The DMA can support any number of features used to provide optimizations for the PVA, including but not limited to support for multi-dimensional addressing and / or circular addressing. In some examples, the DMA can support addressing in up to six or more dimensions, which can include block width, block height, block depth, horizontal block stepping, vertical block stepping, and / or depth stepping.

[0118] A vector processor can be a programmable processor that can be designed to efficiently and flexibly execute programming for computer vision algorithms and provide signal processing capabilities. In some examples, the PVA can include a PVA core and two vector processing subsystem partitions. The PVA core can include a processor subsystem, one or more DMA engines (e.g., two DMA engines), and / or other peripherals. The vector processing subsystem can operate as the main processing engine of the PVA and can include a vector processing unit (VPU), an instruction cache, and / or a vector memory (e.g., VMEM). The VPU core can include a digital signal processor, such as, for example, a single instruction multiple data (SIMD), a very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW can enhance throughput and speed.

[0119] Each of the vector processors can include an instruction cache and can be coupled to dedicated memory. As a result, in some examples, each of the vector processors can be configured to execute independently of the other vector processors. In other examples, the vector processors included in a particular PVA can be configured to employ data parallelism. For example, in some embodiments, multiple vector processors included in a single PVA can execute the same computer vision algorithm, but on different regions of an image. In other examples, the vector processors included in a particular PVA can execute different computer vision algorithms simultaneously on the same image, or even execute different algorithms on sequential images or portions of images. Among other things, any number of PVAs can be included in a hardware accelerator cluster, and any number of vector processors can be included in each of these PVAs. In addition, the PVAs can include additional error correction code (ECC) memory to enhance overall system security.

[0120] The accelerator 814 (e.g., a hardware accelerator cluster) may include an on-chip computer vision network and SRAM to provide high bandwidth, low latency SRAM for the accelerator 814. In some examples, the on-chip memory may include at least 4MB of SRAM consisting of, for example and without limitation, eight field-configurable memory blocks that can be accessed by both the PVA and the DLA. Each pair of memory blocks may include an advanced peripheral bus (APB) interface, configuration circuitry, a controller, and a multiplexer. Any type of memory may be used. The PVA and DLA may access the memory via a backbone that provides high-speed memory access to the PVA and DLA. The backbone may include an on-chip computer vision network that interconnects the PVA and DLA to the memory (e.g., using APB).

[0121] The on-chip computer vision network can include an interface that ensures that both the PVA and DLA provide ready and valid signals before transmitting any control signals / addresses / data. Such an interface can provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst-based communication for continuous data transmission. This type of interface can comply with ISO 26262 or IEC 61508 standards, but other standards and protocols can also be used.

[0122] In some examples, the SoC 804 may include a real-time ray tracing hardware accelerator such as that described in U.S. patent application Ser. No. 16 / 101,232 filed on Aug. 10, 2018. The real-time ray tracing hardware accelerator may be used to quickly and efficiently determine the position and extent of objects (e.g., within a world model) to generate real-time visualization simulations for use in RADAR signal interpretation, for sound propagation synthesis and / or analysis, for SONAR system simulations, for general wave propagation simulations, for comparison with LIDAR data for positioning and / or other functions, and / or for other uses. In some embodiments, one or more tree traversal units (TTUs) may be used to perform one or more ray tracing related operations.

[0123] The accelerator 814 (e.g., a hardware accelerator cluster) has a wide range of uses in autonomous driving. The PVA can be a programmable vision accelerator that can be used in key processing stages in ADAS and autonomous vehicles. The capabilities of the PVA are a good match for algorithmic domains that require predictable processing, low power, and low latency. In other words, the PVA performs well on semi-intensive or intensive rule computations, and even on small data sets that require predictable runtimes with low latency and low power. Therefore, in the context of a platform for autonomous vehicles, the PVA is designed to run classic computer vision algorithms because they are efficient at object detection and integer math operations.

[0124] For example, according to one embodiment of the technology, PVA is used to perform computer stereo vision. In some examples, a semi-global matching-based algorithm can be used, but this is not intended to be limiting. Many applications for Level 3-5 autonomous driving require on-the-fly motion estimation / stereo matching (e.g., structure from motion, pedestrian recognition, lane detection, etc.). PVA can perform computer stereo vision functions on input from two monocular cameras.

[0125] In some examples, PVA can 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 examples, PVA is used for time-of-flight depth processing, by processing raw time-of-flight data to provide processed time-of-flight data.

[0126] The DLA can be used to run any type of network to enhance control and driving safety, including, for example, a neural network that outputs a confidence measure for each object detection. Such confidence values can be interpreted as probabilities, or as providing a relative "weight" of each detection compared to other detections. This confidence value enables the system to make further decisions about which detections should be considered true positives versus false positives. For example, the system can set a threshold for confidence and only consider detections that exceed the threshold as true positives. In an automatic emergency braking (AEB) system, a false positive detection could cause the vehicle to automatically apply emergency braking, which is clearly undesirable. Therefore, only the most confident detections should be considered triggers for AEB. The DLA can run a neural network to regress the confidence value. This neural network can take as its input at least some subset of parameters, such as bounding box dimensions, a ground plane estimate obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 866 output related to the vehicle 800's orientation and distance, and 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., LIDAR sensor 864 or RADAR sensor 860).

[0127] The SoC 804 may include one or more data stores 816 (e.g., memory). The data store 816 may be on-chip memory of the SoC 804 that may store neural networks to be executed on the GPU and / or DLA. In some examples, the data store 816 may be large enough to store multiple instances of the neural network for redundancy and safety. The data store 812 may include an L2 or L3 cache 812. References to the data store 816 may include references to memory associated with the PVA, DLA, and / or other accelerators 814 as described herein.

[0128] The SoC 804 may include one or more processors 810 (e.g., embedded processors). The processors 810 may include a boot and power management processor, which may be a dedicated processor and subsystem for handling boot power and management functions and related safety implementations. The boot and power management processor may be part of the SoC 804 boot sequence and may provide runtime power management services. The boot power and management processor may provide clock and voltage programming, auxiliary system low power state transitions, SoC 804 thermal and temperature sensor management, and / or SoC 804 power state management. Each temperature sensor may be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 804 may use the ring oscillator to detect the temperature of the CPU 806, GPU 808, and / or accelerator 814. If it is determined that the temperature exceeds a threshold, the boot and power management processor may enter a temperature fault routine and place the SoC 804 in a lower power state and / or place the vehicle 800 in a driver safety parking mode (e.g., to safely park the vehicle 800).

[0129] The processor 810 may also include a set of embedded processors that can be used as an audio processing engine. The audio processing engine can be an audio subsystem that allows for full hardware support for multi-channel audio through multiple interfaces and a wide range of flexible audio I / O interfaces. In some examples, the audio processing engine is a dedicated processor core having a digital signal processor with dedicated RAM.

[0130] The processor 810 may also include an always-on processor engine that can provide the necessary hardware features to support low-power sensor management and wake-up use cases. The always-on processor engine may include a processor core, tightly coupled RAM, supporting peripherals (such as timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0131] The processor 810 may also include a safety cluster engine, which includes a dedicated processor subsystem that handles safety management of automotive applications. The safety cluster engine may include two or more processor cores, tightly coupled RAM, supporting peripherals (such as timers, interrupt controllers, etc.), and / or routing logic. In safety mode, the two or more cores can operate in lockstep mode and act as a single core with comparison logic to detect any differences between their operations.

[0132] Processor 810 may also include a real-time camera engine, which may include a dedicated processor subsystem for handling real-time camera management.

[0133] Processor 810 may also include a high dynamic range signal processor, which may include an image signal processor, which is a hardware engine that is part of the camera processing pipeline.

[0134] The processor 810 may include a video image compositer, which may be a processing block (e.g., implemented on a microprocessor) that implements the video post-processing functions required by the video playback application to produce the final image for the player window. The video image compositer may perform lens distortion correction for the wide-angle camera 870, the surround camera 874, and / or for the in-cab monitoring camera sensor. The in-cab monitoring camera sensor is preferably monitored by a neural network running on another instance of the advanced SoC, configured to recognize in-cab events and respond accordingly. The in-cab system may perform lip reading to activate mobile phone service and place calls, dictate emails, change vehicle destinations, activate or change the vehicle's infotainment system and settings, or provide voice-activated web surfing. Certain functions are only available to the driver when the vehicle is operating in autonomous mode and are disabled in other circumstances.

[0135] The video image compositer can include enhanced temporal noise reduction for both spatial and temporal noise reduction. For example, in the presence of motion in the video, the noise reduction appropriately weights spatial information and downweights information provided by neighboring frames. In the case where an image or portion of an image does not include motion, the temporal noise reduction performed by the video image compositer can use information from previous images to reduce noise in the current image.

[0136] The video image compositor can also be configured to perform stereo rectification on the input stereo footage frames. The video image compositor can further be used for user interface composition when the operating system desktop is in use and the GPU 808 does not need to continuously render new surfaces. Even when the GPU 808 is powered on and active for 3D rendering, the video image compositor can be used to offload the GPU 808 to improve performance and responsiveness.

[0137] The SoC 804 may also 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 SoC 804 may also include an input / output controller that may be controlled by software and may be used to receive I / O signals that are not assigned to a specific role.

[0138] The SoC 804 may also include a wide range of peripheral interfaces to enable communication with peripherals, audio codecs, power management, and / or other devices. The SoC 804 may be used to process data from cameras (connected via Gigabit multimedia serial links and Ethernet), sensors (e.g., LIDAR sensor 864, RADAR sensor 860, etc., which may be connected via Ethernet), data from the bus 802 (e.g., vehicle 800 speed, steering wheel position, etc.), and data from the GNSS sensor 858 (connected via Ethernet or a CAN bus). The SoC 804 may also include dedicated high-performance mass storage controllers, which may include their own DMA engines and which may be used to free the CPU 806 from routine data management tasks.

[0139] SoC 804 can be an end-to-end platform with a flexible architecture that spans levels 3-5 of automation, providing a comprehensive functional safety architecture that leverages and efficiently uses computer vision and ADAS technologies for diversity and redundancy, along with deep learning tools to provide a platform for a flexible and reliable driving software stack. SoC 804 can be faster, more reliable, and even more energy- and space-efficient than conventional systems. For example, when combined with CPU 806, GPU 808, and data storage 816, accelerator 814 can provide a fast and efficient platform for level 3-5 autonomous vehicles.

[0140] This technology therefore provides capabilities and functionality that cannot be achieved with conventional systems. For example, computer vision algorithms can be executed on CPUs, which can be configured using high-level programming languages such as the C programming language to execute a wide variety of processing algorithms across a wide variety of visual data. However, CPUs often fail to meet the performance requirements of many computer vision applications, such as those related to execution time and power consumption. In particular, many CPUs are unable to execute complex object detection algorithms in real time, a requirement for in-vehicle ADAS applications and practical Level 3-5 autonomous vehicles.

[0141] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a cluster of hardware accelerators, the technology described herein allows multiple neural networks to be executed simultaneously and / or sequentially, and the results to be combined to achieve Level 3-5 autonomous driving capabilities. For example, a CNN executed on a DLA or dGPU (e.g., GPU 820) can include text and word recognition, allowing the supercomputer to read and understand traffic signs, including signs for which a neural network has not been specifically trained. The DLA can also include a neural network that can recognize, interpret, and provide semantic understanding of the signs, and pass that semantic understanding to a path planning module running on the CPU complex.

[0142] As another example, as required for Level 3, 4, or 5 driving, multiple neural networks can be run simultaneously. For example, a warning sign consisting of "Caution: Flashing lights indicate icing conditions" along with a light can be interpreted by several neural networks, either independently or collectively. The sign itself can be identified as a traffic sign by a first neural network deployed (e.g., a trained neural network), and the text "Flashing lights indicate icing conditions" can be interpreted by a second neural network deployed, which informs the vehicle's path planning software (preferably executing on a CPU complex) that icing conditions exist when the flashing lights are detected. The flashing lights can be identified by operating a third neural network deployed over multiple frames, which informs the vehicle's path planning software of the presence (or absence) of the flashing lights. All three neural networks can run simultaneously, for example, within the DLA and / or on GPU 808.

[0143] In some examples, a CNN for facial recognition and owner recognition can use data from a camera sensor to identify the presence of an authorized driver and / or owner of the vehicle 800. 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, disable the vehicle when the owner leaves the vehicle. In this way, the SoC 804 provides security against theft and / or carjacking.

[0144] In another example, a CNN for emergency vehicle detection and identification can use data from microphone 896 to detect and identify emergency vehicle sirens. In contrast to conventional systems that use general classifiers to detect sirens and manually extract features, SoC 804 uses CNN to classify environmental and urban sounds and to classify visual data. In a preferred embodiment, the CNN running on the DLA is trained to identify the relative closing speed of emergency vehicles (for example, by using the Doppler effect). The CNN can also be trained to identify emergency vehicles specific to the local area in which the vehicle is operating, as identified by the GNSS sensor 858. Thus, for example, when operating in Europe, the CNN will seek to detect European sirens, and when in the United States, the CNN will seek to identify only North American sirens. Once an emergency vehicle is detected, with the assistance of the ultrasonic sensor 862, the control program can be used to execute the emergency vehicle safety routine to slow the vehicle, pull over, stop the vehicle, and / or idle the vehicle until the emergency vehicle passes.

[0145] The vehicle may include a CPU 818 (e.g., a discrete CPU or dCPU) that may be coupled to the SoC 804 via a high-speed interconnect (e.g., PCIe). The CPU 818 may include, for example, an X86 processor. The CPU 818 may be used to perform any of a variety of functions, including, for example, arbitrating potentially inconsistent results between ADAS sensors and the SoC 804, and / or monitoring the status and health of the controller 836 and / or the infotainment SoC 830.

[0146] The vehicle 800 may include a GPU 820 (e.g., a discrete GPU or dGPU) that may be coupled to the SoC 804 via a high-speed interconnect (e.g., NVIDIA's NVLINK). The GPU 820 may provide additional artificial intelligence functionality, for example, by executing redundant and / or different neural networks, and may be used to train and / or update the neural network based at least in part on input from sensors of the vehicle 800 (e.g., sensor data).

[0147] The vehicle 800 may also include a network interface 824, which may include one or more wireless antennas 826 (e.g., one or more wireless antennas for different communication protocols, such as a cellular antenna, a Bluetooth antenna, etc.). The network interface 824 can be used to enable wireless connections to the cloud (e.g., to a server 878 and / or other network devices), to other vehicles, and / or to computing devices (e.g., a passenger's client device) via the Internet. In order to communicate with other vehicles, a direct link can be established between the two vehicles, and / or an indirect link can be established (e.g., across a network and through the Internet). The direct link can be provided using a vehicle-to-vehicle communication link. The vehicle-to-vehicle communication link can provide the vehicle 800 with information about vehicles approaching the vehicle 800 (e.g., vehicles in front of, to the side of, and / or behind the vehicle 800). This functionality can be part of the cooperative adaptive cruise control functionality of the vehicle 800.

[0148] The network interface 824 may include a SoC that provides modulation and demodulation functions and enables the controller 836 to communicate over a wireless network. The network interface 824 may include an RF front-end for up-conversion from baseband to radio frequency and down-conversion from radio frequency to baseband. The frequency conversion may be performed by a well-known process and / or may be performed using a super-heterodyne process. In some examples, the RF front-end function may be provided by a separate chip. The network interface may include wireless functions for communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0149] The vehicle 800 may also include data storage 828, which may include off-chip storage (e.g., outside the SoC 804). The data storage 828 may include one or more storage elements, including RAM, SRAM, DRAM, VRAM, flash memory, a hard disk, and / or other components and / or devices that can store at least one bit of data.

[0150] The vehicle 800 may also include a GNSS sensor 858. The GNSS sensor 858 (e.g., GPS, assisted GPS sensor, differential GPS (DGPS) sensor, etc.) is used to assist with mapping, perception, occupancy grid generation, and / or path planning functions. Any number of GNSS sensors 858 may be used, including, for example and without limitation, GPS using a USB connector with an Ethernet to serial (RS-232) bridge.

[0151] The vehicle 800 may also include a RADAR sensor 860. The RADAR sensor 860 can be used by the vehicle 800 for remote vehicle detection even in darkness and / or in adverse weather conditions. The RADAR functional safety level can be ASILB. The RADAR sensor 860 can use CAN and / or bus 802 (e.g., to transmit data generated by the RADAR sensor 860) for control and access to object tracking data, and in some examples access Ethernet to access raw data. A variety of RADAR sensor types can be used. For example and without limitation, the RADAR sensor 860 can be suitable for front, rear, and side RADAR use. In some examples, a pulse Doppler RADAR sensor is used.

[0152] The RADAR sensor 860 can include different configurations, such as long-range with a narrow field of view, short-range with a wide field of view, short-range side coverage, and so on. In some examples, the long-range RADAR can be used for adaptive cruise control functions. The long-range RADAR system can provide a wide field of view (e.g., within a range of 250m) achieved by two or more independent scans. The RADAR sensor 860 can help distinguish between static objects and moving objects and can be used by the ADAS system for emergency braking assistance and forward collision warnings. The long-range RADAR sensor may include a single-station multimode RADAR with multiple (e.g., six or more) fixed RADAR antennas and high-speed CAN and FlexRay interfaces. In the example with six antennas, the central four antennas can create a focused beam pattern that is designed to record the surroundings of the vehicle 800 at a higher rate with minimal traffic interference from adjacent lanes. The other two antennas can expand the field of view, making it possible to quickly detect vehicles entering or leaving the lane of the vehicle 800.

[0153] As an example, a medium-range RADAR system may include a range of up to 860m (front) or 80m (rear) and a field of view of up to 42 degrees (front) or 850 degrees (rear). A short-range RADAR system may include, but is not limited to, a RADAR sensor designed to be mounted on both ends of the rear bumper. When mounted on both ends of the rear bumper, such a RADAR sensor system can create two beams that continuously monitor the blind spots behind and beside the vehicle.

[0154] Short-range RADAR systems can be used in ADAS systems for blind spot detection and / or lane change assistance.

[0155] Vehicle 800 may also include ultrasonic sensors 862. Ultrasonic sensors 862, which may be located on the front, rear, and / or sides of vehicle 800, may be used for parking assistance and / or for creating and updating an occupancy grid. A variety of ultrasonic sensors 862 may be used, and different ultrasonic sensors 862 may have different detection ranges (e.g., 2.5 m, 4 m). Ultrasonic sensors 862 may operate at functional safety level ASIL B.

[0156] Vehicle 800 may include a LIDAR sensor 864. LIDAR sensor 864 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. LIDAR sensor 864 may be ASIL B functional safety level. In some examples, vehicle 800 may include multiple LIDAR sensors 864 (e.g., two, four, six, etc.) that may use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0157] In some examples, the LIDAR sensor 864 may be capable of providing a list of objects and their distances for a 360-degree field of view. Commercially available LIDAR sensors 864 may have, for example, an advertised range of approximately 800 meters, an accuracy of 2-3 cm, and support for 800 Mbps Ethernet connections. In some examples, one or more non-obtrusive LIDAR sensors 864 may be used. In such examples, the LIDAR sensor 864 may be implemented as a small device that can be embedded in the front, back, sides, and / or corners of the vehicle 800. In such examples, the LIDAR sensor 864 may provide a field of view of up to 120 degrees horizontally and 35 degrees vertically, with a range of 200 meters, even for low-reflectivity objects. The front-mounted LIDAR sensor 864 may be configured for a horizontal field of view between 45 and 135 degrees.

[0158] In some examples, LIDAR technologies such as 3D flash LIDAR may also be used. 3D flash LIDAR uses flashes of laser as an emission source to illuminate the vehicle's surroundings up to about 200 m. The flash LIDAR unit includes a receiver that records the laser pulse transmission time and reflected light on each pixel, which in turn corresponds to the range from the vehicle to the object. Flash LIDAR can allow a highly accurate and distortion-free image of the surrounding environment to be generated with each laser flash. In some examples, four flash LIDAR sensors can be deployed, one on each side of the vehicle 800. Available 3D flash LIDAR systems include solid-state 3D staring array LIDAR cameras (e.g., non-scanning LIDAR devices) with no moving parts other than a fan. The flash LIDAR device can use 5 nanosecond Class I (eye-safe) laser pulses per frame and can capture the reflected laser light in the form of a 3D range point cloud and co-registered intensity data. By using flash LIDAR, and because flash LIDAR is a solid-state device with no moving parts, the LIDAR sensor 864 may be less susceptible to motion blur, vibration, and / or shock.

[0159] The vehicle may also include an IMU sensor 866. In some examples, the IMU sensor 866 may be located at the center of the rear axle of the vehicle 800. The IMU sensor 866 may include, for example and without limitation, an accelerometer, a magnetometer, a gyroscope, a magnetic compass, and / or other sensor types. In some examples, such as in a six-axis application, the IMU sensor 866 may include an accelerometer and a gyroscope, while in a nine-axis application, the IMU sensor 866 may include an accelerometer, a gyroscope, and a magnetometer.

[0160] In some embodiments, the IMU sensor 866 can be implemented as a miniature high-performance GPS-assisted inertial navigation system (GPS / INS) that combines micro-electromechanical systems (MEMS) inertial sensors, a high-sensitivity GPS receiver, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. Thus, in some examples, the IMU sensor 866 can enable the vehicle 800 to estimate heading without the need for input from a magnetic sensor by directly observing and correlating velocity changes from the GPS to the IMU sensor 866. In some examples, the IMU sensor 866 and the GNSS sensor 858 can be combined into a single integrated unit.

[0161] The vehicle may include microphones 896 positioned in and / or around the vehicle 800. The microphones 896 may be used for, among other things, emergency vehicle detection and identification.

[0162] The vehicle may also include any number of camera types, including stereo cameras 868, wide angle cameras 870, infrared cameras 872, surround cameras 874, long and / or medium range cameras 898, and / or other camera types. These cameras may be used to capture image data around the entire periphery of the vehicle 800. The type of camera used depends on the embodiment and the requirements of the vehicle 800, and any combination of camera types may be used to provide the necessary coverage around the vehicle 800. 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. As an example and not limitation, the cameras may support Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Each of the cameras described herein may include a GMSL and / or Gigabit Ethernet network. Figure 8A and Figure 8B Described in more detail.

[0163] Vehicle 800 may also include a vibration sensor 842. Vibration sensor 842 can measure vibrations of vehicle components, such as axles. For example, changes in vibration can indicate changes in the road surface. In another example, when two or more vibration sensors 842 are used, the difference between the vibrations can be used to determine friction or slippage of the road surface (e.g., when there is a vibration difference between a powered drive shaft and a freely rotating shaft).

[0164] The vehicle 800 may include an ADAS system 838. In some examples, the ADAS system 838 may include a SoC. The ADAS system 838 may include autonomous / adaptive / automatic cruise control (ACC), cooperative adaptive cruise control (CACC), forward collision warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keeping assist (LKA), blind spot warning (BSW), rear cross traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functions.

[0165] The ACC system can utilize RADAR sensor 860, LIDAR sensor 864, and / or cameras. The ACC system can include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle immediately in front of vehicle 800, automatically adjusting the vehicle speed to maintain a safe distance from the vehicle in front. Lateral ACC maintains distance and, when necessary, recommends that vehicle 800 change lanes. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0166] CACC uses information from other vehicles, which can be received from other vehicles indirectly via a wireless link or through a network connection (e.g., through the Internet) via the network interface 824 and / or the wireless antenna 826. A direct link can be provided by a vehicle-to-vehicle (V2V) communication link, while an indirect link can be an infrastructure-to-vehicle (I2V) communication link. Typically, the V2V communication concept provides information about the vehicle immediately ahead (e.g., the vehicle immediately ahead of the vehicle 800 and in the same lane as it), while the I2V communication concept provides information about traffic further ahead. The CACC system can include either or both of the I2V and V2V information sources. Given information about the vehicle ahead of the vehicle 800, CACC can be more reliable, and it has the potential to improve the smoothness of traffic flow and reduce road congestion.

[0167] The FCW system is designed to alert the driver to hazards so that the driver can take corrective action. The FCW system uses a front-facing camera and / or RADAR sensor 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC, which is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component. The FCW system can provide warnings in the form of, for example, audible, visual warnings, vibrations, and / or rapid brake pulses.

[0168] The AEB system detects an impending forward collision with another vehicle or other object and can automatically apply the brakes if the driver does not take corrective action within specified time or distance parameters. The AEB system can use a front-facing camera and / or RADAR sensor 860 coupled to a dedicated processor, DSP, FPGA and / or ASIC. When the AEB system detects a hazard, it 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 in an effort to prevent or at least mitigate the effects of the predicted collision. The AEB system can include technologies such as dynamic brake support and / or collision approach braking.

[0169] The LDW system provides visual, audible, and / or tactile warnings, such as steering wheel or seat vibrations, to alert the driver when the vehicle 800 crosses a lane marking. When the driver indicates an intention to leave the lane by activating a turn signal, the LDW system is deactivated. The LDW system may utilize a front-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibration component.

[0170] The LKA system is a variation of the LDW system. If the vehicle 800 begins to leave its lane, the LKA system provides steering input or braking to correct the vehicle 800.

[0171] The BSW system detects and warns the driver of vehicles in the car's blind spot. The BSW system can provide visual, audible, and / or tactile alerts to indicate that merging or changing lanes is unsafe. The system can provide additional warnings when the driver uses a turn signal. The BSW system can use a rear-facing camera and / or RADAR sensor 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibration component.

[0172] The RCTW system can provide visual, audible, and / or tactile notifications when an object is detected outside the range of the rear-mounted camera while the vehicle 800 is in reverse. Some RCTW systems include automatic emergency braking (AEB) to ensure that the vehicle brakes are applied to avoid a collision. The RCTW system can use one or more rear-mounted RADAR sensors 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibration component.

[0173] Conventional ADAS systems may be prone to false positive results, which may be annoying and distracting to the driver, but are typically not catastrophic because the ADAS system alerts the driver and allows the driver to decide whether a safety condition actually exists and take action accordingly. However, in the autonomous vehicle 800, in the event of conflicting results, the vehicle 800 itself must decide whether to pay attention to the results from the main computer or the auxiliary computer (e.g., the first controller 836 or the second controller 836). For example, in some embodiments, the ADAS system 838 can be a backup and / or auxiliary computer for providing perception information to the backup computer rationality module. The backup computer rationality monitor can run redundant and diverse software on hardware components to detect failures in perception and dynamic driving tasks. The output from the ADAS system 838 can be provided to the supervisory MCU. If the outputs from the main computer and the auxiliary computer conflict, the supervisory MCU must determine how to reconcile the conflict to ensure safe operation.

[0174] In some examples, the primary computer can be configured to provide a confidence score to the supervisory MCU, indicating the primary computer's confidence in the selected result. If the confidence score exceeds a threshold, the supervisory MCU can follow the primary computer's direction, regardless of whether the secondary computer provides conflicting or inconsistent results. In the event that the confidence score does not meet the threshold and the primary and secondary computers indicate different results (e.g., a conflict), the supervisory MCU can arbitrate between these computers to determine the appropriate result.

[0175] The supervisory MCU can be configured to run a neural network that is trained and configured to determine conditions under which the secondary computer provides a false alarm based, at least in part, on outputs from the primary computer and the secondary computer. Thus, the neural network in the supervisory MCU can learn when the output of the secondary computer can be trusted and when it cannot. For example, when the secondary computer is a RADAR-based FCW system, the neural network in the supervisory MCU can learn when the FCW system is identifying a metal object that is not actually a danger, such as a drain grate or manhole cover, which triggers an alarm. Similarly, when the secondary computer is a camera-based LDW system, the neural network in the supervisory MCU can learn to disregard the LDW when a cyclist or pedestrian is present and lane departure is actually the safest strategy. In embodiments that include a neural network running on the supervisory MCU, the supervisory MCU can include at least one of a DLA or a GPU suitable for running the neural network with associated memory. In a preferred embodiment, the supervisory MCU can include and / or be included as a component of the SoC 804.

[0176] In other examples, the ADAS system 838 may include an auxiliary computer that uses traditional computer vision rules to perform ADAS functions. In this way, the auxiliary computer can use classic computer vision rules (if-then), and the presence of a neural network in the supervisory MCU can improve reliability, safety, and performance. For example, diverse implementations and intentional non-identity make the entire system more fault-tolerant, especially with respect to failures caused by software (or software-hardware interface) functions. For example, if there is a software vulnerability or bug in the software running on the main computer and the non-identical software code running on the auxiliary computer provides the same overall result, the supervisory MCU can be more confident that the overall result is correct and that the vulnerability in the software or hardware on the main computer did not cause a substantial error.

[0177] In some examples, the output of the ADAS system 838 can be fed into the primary computer's perception block and / or the primary computer's dynamic driving task block. For example, if the ADAS system 838 indicates a forward collision warning due to an object immediately ahead, the perception block can use this information when identifying the object. In other examples, the secondary computer can have its own neural network that is trained and thus reduces the risk of false positives as described herein.

[0178] The vehicle 800 may also include an infotainment SoC 830 (e.g., an in-vehicle infotainment system (IVI)). Although illustrated and described as an SoC, the infotainment system may not be an SoC and may include two or more separate components. The infotainment SoC 830 may include a combination of hardware and software that can be used to provide audio (e.g., music, personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., TV, movies, streaming, etc.), phone (e.g., hands-free calling), network connectivity (e.g., LTE, WiFi, etc.), and / or information services (e.g., a navigation system, rear parking assistance, radio data system, vehicle-related information such as fuel level, total distance covered, brake fuel level, oil level, door open / closed, air filter information, etc.) to the vehicle 800. For example, the infotainment SoC 830 may include a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, an onboard computer, in-vehicle entertainment, WiFi, steering wheel audio controls, hands-free voice controls, a head-up display (HUD), an HMI display 834, 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 830 may further be used to provide information (e.g., visual and / or auditory) to a user of the vehicle, such as information from an ADAS system 838, autonomous driving information such as planned vehicle maneuvers, trajectories, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0179] The infotainment SoC 830 may include GPU functionality. The infotainment SoC 830 may communicate with other devices, systems, and / or components of the vehicle 800 via a bus 802 (e.g., a CAN bus, Ethernet, etc.). In some examples, the infotainment SoC 830 may be coupled to a supervisory MCU so that in the event of a failure of a primary controller 836 (e.g., a primary and / or backup computer of the vehicle 800), the infotainment system's GPU may perform some self-driving functions. In such an example, the infotainment SoC 830 may place the vehicle 800 in a driver-safe parking mode as described herein.

[0180] The vehicle 800 may also include an instrument cluster 832 (e.g., a digital instrument panel, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 832 may include a controller and / or a supercomputer (e.g., a separate controller or a supercomputer). The instrument cluster 832 may include a set of instruments, such as a speedometer, fuel level, oil pressure, a tachometer, an odometer, a turn indicator, a shift position indicator, a seat belt warning light, a parking brake warning light, an engine check light, airbag (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some examples, information can be displayed and / or shared between the infotainment SoC 830 and the instrument cluster 832. In other words, the instrument cluster 832 may be included as part of the infotainment SoC 830, or vice versa.

[0181] Figure 8D For cloud-based servers and Figure 8A 880). Depending on the embodiment, each of the servers 878 can include any number of GPUs 884, CPUs 880, and / or PCIe switches. For example, each of the servers 878 can include eight, sixteen, thirty-two, and / or more GPUs 884.

[0182] Server 878 can receive image data from a vehicle via network 890 that represents images showing unexpected or changed road conditions, such as recently begun road construction. Server 878 can transmit neural network 892, updated neural network 892, and / or map information 894, including information about traffic and road conditions, via network 890 and to the vehicle. Updates to map information 894 can include updates to HD map 822, such as information about construction sites, potholes, curves, flooding, or other obstacles. In some examples, neural network 892, updated neural network 892, and / or map information 894 can have been generated from new training and / or data received from any number of vehicles in the environment and / or based on experience with training performed at a data center (e.g., using server 878 and / or other servers).

[0183] Server 878 can be used to train a machine learning model (e.g., a neural network) based on training data. The training data can be generated by the vehicle and / or can be generated in simulation (e.g., using a game engine). In some examples, the training data is labeled (e.g., in cases where the neural network benefits from supervised learning) and / or undergoes other preprocessing, while in other examples, the training data is not labeled and / or preprocessed (e.g., in cases where the neural network does not require supervised learning). Training can be performed according to any one or more categories of machine learning techniques, including but not limited to the following categories: supervised training, semi-supervised training, unsupervised training, self-learning, reinforcement learning, joint learning, transfer learning, feature learning (including principal component and cluster analysis), multilinear subspace learning, manifold learning, representation learning (including alternative dictionary learning), rule-based machine learning, anomaly detection, and any variants or combinations thereof. Once the machine learning model is trained, the machine learning model can be used by the vehicle (e.g., transmitted to the vehicle via network 890), and / or the machine learning model can be used by server 878 to remotely monitor the vehicle.

[0184] In some examples, server 878 can receive data from the vehicle and apply the data to the latest real-time neural network for real-time intelligent reasoning. Server 878 can include a deep learning supercomputer and / or a dedicated AI computer powered by GPU 884, such as the DGX and DGX Station machines developed by NVIDIA. However, in some examples, server 878 can include the deep learning infrastructure of a data center using only CPU power.

[0185] The deep learning infrastructure of server 878 may be capable of rapid real-time inference and may use this capability to assess and verify the health of the processors, software, and / or associated hardware in vehicle 800. For example, the deep learning infrastructure may receive periodic updates from vehicle 800, such as an image sequence and / or objects that vehicle 800 has located in the image 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 to the objects identified by vehicle 800, and if the results do not match and the infrastructure concludes that the AI in vehicle 800 has malfunctioned, server 878 may transmit a signal to vehicle 800 instructing the vehicle's 800 fail-safe computer to take control, notify passengers, and complete a safe parking maneuver.

[0186] For inference, the server 878 may include a GPU 884 and one or more programmable inference accelerators (e.g., NVIDIA's TensorRT). The combination of GPU-powered servers and inference acceleration can enable real-time responses. In other examples, such as where performance is less important, CPU, FPGA, and other processor-powered servers can be used for inference.

[0187] Example computing device

[0188] Figure 9 9 is a block diagram of an example computing device 900 suitable for implementing some embodiments of the present disclosure. Computing device 900 may include an interconnect system 902 that directly or indirectly couples the following devices: memory 904, one or more central processing units (CPUs) 906, one or more graphics processing units (GPUs) 908, a communication interface 910, input / output (I / O) ports 912, input / output components 914, a power supply 916, one or more presentation components 918 (e.g., display(s)), and one or more logic units 920. In at least one embodiment, computing device(s) 900 may include one or more virtual machines (VMs), and / or any of its components may include virtual components (e.g., virtual hardware components). For non-limiting examples, one or more of GPUs 908 may include one or more vGPUs, one or more of CPUs 906 may include one or more vCPUs, and / or one or more of logic units 920 may include one or more virtual logic units. As such, computing device(s) 900 may include discrete components (e.g., a full GPU dedicated to computing device 900), virtual components (e.g., a portion of a GPU dedicated to computing device 900), or a combination thereof.

[0189] although Figure 9 The various blocks of are shown as being connected via interconnect system 902 using wires, but this is not intended to be limiting and is provided for clarity only. For example, in some embodiments, presentation component 918 (such as a display device) may be considered to be I / O component 914 (e.g., if the display is a touch screen). As another example, CPU 906 and / or GPU 908 may include memory (e.g., memory 904 may represent a storage device in addition to the memory of GPU 908, CPU 906, and / or other components). In other words, Figure 9 The computing devices are illustrative only. No distinction is made between such categories as "workstation," "server," "laptop," "desktop," "tablet," "client device," "mobile device," "handheld device," "game console," "electronic control unit (ECU)," "virtual reality system," and / or other device or system types, as all are considered Figure 9 within the range of computing devices.

[0190] Interconnect system 902 can represent one or more links or buses, such as address bus, data bus, control bus or its combination.Interconnect system 902 can include one or more bus or link types, such as industry standard architecture (ISA) bus, extended industry standard architecture (EISA) bus, video electronics standard association (VESA) bus, peripheral component interconnect (PCI) bus, peripheral component interconnect express (PCIe) bus and / or another type of bus or link. In some embodiments, there is a direct connection between components. As an example, CPU 906 can be directly connected to memory 904. Further, CPU 906 can be directly connected to GPU 908. In the case of direct or point-to-point connection between components, interconnect system 902 can include PCIe link to perform connection. In these examples, PCI bus does not need to be included in computing device 900.

[0191] Memory 904 may include any of a variety of computer-readable media. Computer-readable media can be any available media that can be accessed by computing device 900. Computer-readable media can include volatile and non-volatile media, as well as removable and non-removable media. By way of example and not limitation, computer-readable media can include computer storage media and communication media.

[0192] Computer storage media may include volatile and nonvolatile media and / or removable and non-removable media implemented in any method or technology for storage of information such as computer-readable instructions, data structures, program modules, and / or other data types. For example, memory 904 may store computer-readable instructions (e.g., representing (one or more) programs and / or (one or more) program elements, such as an operating system). Computer storage media may include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage devices or other magnetic storage devices, or any other medium that can be used to store the desired information and that can be accessed by the computing device 900. As used herein, computer storage media does not include signals themselves.

[0193] Computer storage media can 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, computer storage media can 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.

[0194] The CPU 906 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. The CPUs 906 may each include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of handling numerous software threads simultaneously. The CPU 906 may include any type of processor and may include different types of processors depending on the type of computing device 900 being 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 900, the processor may be an Advanced RISC Machine (ARM) processor implemented using Reduced Instruction Set Computing (RISC) or an x86 processor implemented using Complex Instruction Set Computing (CISC). The computing device 900 may also include one or more CPUs 906 in addition to one or more microprocessors or supplemental coprocessors (such as a math coprocessor).

[0195] In addition to or in lieu of CPU(s) 906, GPU(s) 908 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. One or more of the GPUs 908 may be integrated GPUs (e.g., with one or more of the CPUs 906) and / or one or more of the GPUs 908 may be discrete GPUs. In embodiments, one or more of the GPUs 908 may be coprocessors for one or more of the CPUs 906. The GPU 908 may be used by the computing device 900 to render graphics (e.g., 3D graphics) or perform general-purpose computations. For example, the GPU 908 may be used for general-purpose computing on a GPU (GPGPU). The GPU 908 may include hundreds or thousands of cores capable of handling hundreds or thousands of software threads simultaneously. The GPU 908 may generate pixel data for an output image in response to rendering commands (e.g., rendering commands received from the CPU 906 via a host interface). The GPU 908 may include graphics memory (e.g., display memory) for storing pixel data or any other suitable data (e.g., GPGPU data). The display memory may be included as part of the memory 904. The GPU 908 may include two or more GPUs operating in parallel (e.g., via a link). The link may connect the GPUs directly (e.g., using NVLINK) or may connect the GPUs through a switch (e.g., using NVSwitch). When combined, each GPU 908 may generate pixel data or GPGPU data for different portions or for different outputs (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU may include its own memory or may share memory with other GPUs.

[0196] In addition to or in lieu of the CPU 906 and / or GPU 908, the logic unit 920 may be configured to execute at least some of the computer-readable instructions to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. In embodiments, the CPU(s) 906, the GPU(s) 908, and / or the logic unit(s) 920 may perform any combination of methods, processes, and / or portions thereof, either discretely or jointly. One or more of the logic units 920 may be part of and / or integrated into one or more of the CPU 906 and / or GPU 908, and / or one or more of the logic units 920 may be discrete components or otherwise external to the CPU 906 and / or GPU 908. In embodiments, one or more of logic units 920 may be a co-processor of one or more of CPUs 906 and / or one or more of GPUs 908 .

[0197] Examples of logic unit 920 include one or more processing cores and / or components thereof, such as a tensor core (TC), a tensor processing unit (TPU), a pixel vision core (PVC), a vision processing unit (VPU), a data processing unit (DPU), a graphics processing cluster (GPC), a texture processing cluster (TPC), a streaming multiprocessor (SM), a tree transverse unit (TTU), an artificial intelligence accelerator (AIA), a deep learning accelerator (DLA), an arithmetic logic unit (ALU), an application-specific integrated circuit (ASIC), a floating point unit (FPU), an input / output (I / O) element, a peripheral component interconnect (PCI) or a peripheral component interconnect express (PCIe) element, etc.

[0198] The communication interface 910 may include one or more receivers, transmitters, and / or transceivers that enable the computing device 900 to communicate with other computing devices via an electronic communication network (including wired and / or wireless communications). The communication interface 910 may include components and functionality that 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., via Ethernet or Wi-Fi), a low-power wide-area network (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0199] The I / O ports 912 can enable the computing device 900 to be logically coupled to other devices including I / O components 914, (one or more) presentation components 918, and / or other components, some of which can be built into (e.g., integrated into) the computing device 900. Illustrative I / O components 914 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 914 can provide a natural user interface (NUI) that processes air gestures, voice, or other physiological input generated by the user. In some cases, the input can be transmitted to an appropriate network element for further processing. The NUI can implement any combination of voice recognition, stylus recognition, facial recognition, biometric recognition, gesture recognition on and near the screen, air gestures, head and eye tracking, and touch recognition (as described in more detail below) associated with the display of the computing device 900. The computing device 900 may include a depth camera for gesture detection and recognition, such as a stereo camera system, an infrared camera system, an RGB camera system, touch screen technology, and combinations thereof. Additionally, computing device 900 may include an accelerometer or gyroscope that enables detection of motion (e.g., as part of an inertial measurement unit (IMU)). In some examples, computing device 900 may use the output of the accelerometer or gyroscope to render immersive augmented or virtual reality.

[0200] The power supply 916 may include a hardwired power supply, a battery power supply, or a combination thereof. The power supply 916 may provide power to the computing device 900 to enable the components of the computing device 900 to operate.

[0201] The presentation component 918 may include a display (e.g., a monitor, a touch screen, a television screen, a head-up display (HUD), other display types, or a combination thereof), speakers, and / or other presentation components. The presentation component 918 may receive data from other components (e.g., GPU 908, CPU 906, etc.) and output the data (e.g., as images, video, sound, etc.).

[0202] Sample Data Center

[0203] Figure 10 An example data center 1000 that can be used in at least one embodiment of the present disclosure is shown. The data center 1000 can include a data center infrastructure layer 1010, a framework layer 1020, a software layer 1030, and / or an application layer 1040.

[0204] like Figure 10As shown, the data center infrastructure layer 1010 may include a resource coordinator 1012, grouped computing resources 1014, and node computing resources ("node CRs") 1016(1)-1016(N), where "N" represents any complete positive integer. In at least one embodiment, the node CRs 1016(1)-1016(N) may include, but is not limited to, any number of central processing units ("CPUs") or other processors (including accelerators, field programmable gate arrays (FPGAs), graphics processors or graphics processing units (GPUs), etc.), memory devices (e.g., dynamic read-only memories), storage devices (e.g., solid-state or disk drives), network input / output ("NW I / O") devices, network switches, virtual machines ("VMs"), power modules and / or cooling modules, etc. In some embodiments, one or more of the node CRs 1016(1)-1016(N) may correspond to a server having one or more of the above-mentioned computing resources. Furthermore, in some embodiments, node CRs 1016(1)-10161(N) may include one or more virtual components, such as vGPUs, vCPUs, etc., and / or one or more of node CRs 1016(1)-1016(N) may correspond to a virtual machine (VM).

[0205] In at least one embodiment, the grouped computing resources 1014 may include individual groups of node CRs 1016 housed in one or more racks (not shown), or multiple racks housed in data centers at different geographical locations (also not shown). Individual groups of node CRs 1016 within the grouped computing resources 1014 may include grouped computing, network, memory, or storage resources that can be configured or allocated to support one or more workloads. In at least one embodiment, several node CRs 1016 including CPUs, GPUs, and / or other processors may be grouped in one or more racks to provide computing resources to support one or more workloads. One or more racks may also include any number of power modules, cooling modules, and / or network switches in any combination.

[0206] Resource coordinator 1022 may configure or otherwise control one or more node CRs 1016(1)-1016(N) and / or grouped computing resources 1014. In at least one embodiment, resource coordinator 1022 may comprise a software design infrastructure ("SDI") management entity for data center 1000. Resource coordinator 1022 may comprise hardware, software, or some combination thereof.

[0207] In at least one embodiment, Figure 10As shown, the framework layer 1020 may include a job scheduler 1032, a configuration manager 1034, a resource manager 1036, and / or a distributed file system 1038. The framework layer 1020 may include a framework that supports the software 1032 of the software layer 1030 and / or one or more applications 1042 of the application layer 1040. The software 1032 or the application 1042 may include network-based service software or applications, such as those provided by Amazon Web Services, Google Cloud, and Microsoft Azure. The framework layer 1020 may be, but is not limited to, a free and open source software network application framework (such as Apache Spark) that can utilize the distributed file system 1038 for large-scale data processing (e.g., "big data"). TM (hereinafter referred to as "Spark"). In at least one embodiment, the job scheduler 1032 may include a Spark driver to facilitate scheduling workloads supported by the different layers of the data center 1000. The configuration manager 1034 may be capable of configuring the different layers, such as the software layer 1030 and the framework layer 1020 (which includes Spark and a distributed file system 1038 for supporting large-scale data processing). The resource manager 1036 may be capable of managing clustered or grouped computing resources that are mapped to the distributed file system 1038 and the job scheduler 1032 or allocated to support the distributed file system 1038 and the job scheduler 1032. In at least one embodiment, the clustered or grouped computing resources may include the grouped computing resources 1014 at the data center infrastructure layer 1010. The resource manager 1036 may coordinate with the resource coordinator 1012 to manage these mapped or allocated computing resources.

[0208] In at least one embodiment, the software 1032 included in the software layer 1030 may include software used by at least a portion of the node CRs 1016(1)-1016(N), the grouped computing resources 1014, and / or the distributed file system 1038 of the framework layer 1020. The one or more types of software may include, but are not limited to, Internet web search software, email virus scanning software, database software, and streaming video content software.

[0209] In at least one embodiment, the applications 1042 included in the application layer 1040 may include one or more types of applications used by at least a portion of the node CRs 1016(1)-1016(N), the grouped computing resources 1014, and / or the distributed file system 1038 of the framework layer 1020. The one or more types of applications may include, but are not limited to, any number of genomic applications, cognitive computing, and machine learning applications, including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), and / or other machine learning applications used in conjunction with one or more embodiments.

[0210] In at least one embodiment, any of the configuration manager 1034, the resource manager 1036, and the resource coordinator 1012 can implement any number and type of self-modification actions based on any amount and type of data acquired in any technically feasible manner. The self-modification actions can save the data center operator of the data center 1000 from making potentially poor configuration decisions and potentially avoiding underutilized and / or poorly performing portions of the data center.

[0211] According to one or more embodiments described herein, data center 1000 may include tools, services, software, or other resources to train one or more machine learning models or to predict or infer information using one or more machine learning models. For example, the machine learning model(s) may be trained by computing weight parameters according to a neural network architecture using the software and / or computing resources described above with respect to data center 1000. In at least one embodiment, a trained or deployed machine learning model corresponding to one or more neural networks may be used to infer or predict information using the resources described above with respect to data center 1000 using weight parameters computed using one or more training techniques, such as, but not limited to, those described herein.

[0212] In at least one embodiment, data center 1000 may use a CPU, an application-specific integrated circuit (ASIC), a GPU, an FPGA, and / or other hardware (or virtual computing resources corresponding thereto) to perform training and / or inference using the aforementioned resources. In addition, one or more of the software and / or hardware resources described above may be configured to allow a user to train or perform inference services on information, such as image recognition, speech recognition, or other artificial intelligence services.

[0213] Sample network environment

[0214] A network environment suitable for implementing embodiments of the present disclosure may include one or more client devices, servers, network attached storage (NAS), other backend devices, and / or other device types. The client devices, servers, and / or other device types (e.g., each device) may be configured to: Figure 9 The data center 1000 may be implemented on one or more instances of the computing device(s) 900 - for example, each device may include similar components, features, and / or functionality of the computing device(s) 900. In addition, in the case of implementing a backend device (e.g., a server, NAS, etc.), the backend device may be included as part of the data center 1000, an example of which is described herein with respect to FIG. Figure 10 Describe in more detail.

[0215] The components of the network environment can communicate with each other via a network, which can be wired, wireless, or both. The network can include multiple networks or one of multiple networks. For example, the network can include one or more wide area networks (WANs), one or more local area networks (LANs), one or more public networks (such as the Internet and / or the Public Switched Telephone Network (PSTN)), and / or one or more private networks. In the case where the network includes a wireless telecommunications network, components such as base stations, communication towers, or even access points (and other components) can provide wireless connectivity.

[0216] Compatible network environments may include one or more peer-to-peer network environments (in which case the server may not be included in the network environment) and one or more client-server network environments (in which case one or more servers may be included in the network environment). In a peer-to-peer network environment, the functionality described herein for the server may be implemented on any number of client devices.

[0217] In at least one embodiment, the network environment may include one or more cloud-based network environments, distributed computing environments, combinations thereof, and the like. The cloud-based network environment may include a framework layer, a job scheduler, a resource manager, and a distributed file system implemented on one or more servers, which may include one or more core network servers and / or edge servers. The framework layer may include software supporting the software layer and / or a framework for one or more applications at the application layer. The software or application may include network-based service software or applications, respectively. In an embodiment, one or more client devices may use network-based service software or applications (e.g., by accessing the service software and / or application via one or more application programming interfaces (APIs)). The framework layer may be, but is not limited to, a free and open source software network application framework that can use a distributed file system for large-scale data processing (e.g., "big data").

[0218] A cloud-based network environment can provide cloud computing and / or cloud storage that performs any combination of the computing and / or data storage functions described herein (or one or more portions thereof). Any of these different functions can be distributed across multiple locations from a central or core server (e.g., one or more data centers that can be distributed across a state, region, country, global, etc.). If the connection to the user (e.g., client device) is relatively close to an edge server, the core server can assign at least a portion of the functionality to the edge server. A cloud-based network environment can be private (e.g., limited to a single organization), public (e.g., available to many organizations), and / or a combination thereof (e.g., a hybrid cloud environment).

[0219] The client device(s) may include the Figure 9 At least some of the components, features, and functionality of the described example computing device(s) 900. By way of example and not limitation, the client device may be implemented as a personal computer (PC), a laptop computer, a mobile device, a smartphone, a tablet computer, a smartwatch, a wearable computer, a personal digital assistant (PDA), an MP3 player, a virtual reality headset, a global positioning system (GPS) or device, a video player, a camera, a surveillance device or system, a vehicle, a boat, a spacecraft, a virtual machine, a drone, a robot, a handheld communication device, a hospital device, a gaming device or system, an entertainment system, a vehicle computer system, an embedded system controller, a remote control, an appliance, a consumer electronic device, a workstation, an edge device, any combination of the depicted devices, or any other suitable device.

[0220] The present disclosure can be described in the general context of machine-usable instructions or computer code executed by a computer or other machine such as a personal digital assistant or other handheld device, including computer-executable instructions such as program modules. Generally, program modules including routines, programs, objects, components, data structures, etc. refer to code that performs a specific task or implements a specific abstract data type. The present disclosure can be practiced in a variety of system configurations, including handheld devices, consumer electronics, general-purpose computers, more specialized computing devices, etc. The present disclosure can also be practiced in a distributed computing environment where tasks are performed by remote processing devices linked through a communication network.

[0221] As used herein, the phrase "and / or" with respect to two or more elements should be interpreted as referring to only one element or combination of elements. For example, "element A, element B, and / or element C" may include only element A, only element B, only element C, element A and element B, element A and element C, element B and element C, or elements A, B, and C. In addition, "at least one of element A or element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. Further, "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.

[0222] The subject matter of the present disclosure is described in detail herein to meet statutory requirements. However, the description itself is not intended to limit the scope of the present disclosure. On the contrary, the present inventors have contemplated that the claimed subject matter may also be embodied in other ways to include steps that are different from the steps described herein in conjunction with other current or future technologies, or combinations of similar steps. Moreover, although the terms "step" and / or "block" may be used herein to imply different elements of the method employed, these terms should not be interpreted as implying any particular order among or between the various steps disclosed herein, unless the order of the steps is explicitly described.

Claims

1. A processor, comprising: One or more circuits configured to: receive first data representing a plurality of outputs of a plurality of deep neural networks (DNNs), at least one of the plurality of outputs corresponding to a respective sensor having a respective field of view that is different from a field of view corresponding to one or more other sensors of a plurality of sensors of an autonomous machine; and generate, for each of the plurality of outputs, a plurality of rasterized images using the first data, the rasterized images depicting at least one of an object position, an object classification, an object velocity, or an object orientation corresponding to each of one or more objects. computing, using the fused DNN and based at least in part on the plurality of rasterized images, second data representing a fusion of the plurality of rasterized images; and performing one or more operations using the autonomous machine based at least in part on the second data.

2. The processor of claim 1 , wherein the calculation of the second data is further based at least in part on third data, the third data representing at least one probability distribution function corresponding to at least one point of at least one of the plurality of outputs, the at least one point corresponding to a detection object and the at least one probability distribution function corresponding to one or more potential positions of the detection object.

3. A processor according to claim 1, wherein the calculation of the second data is further based at least in part on third data representing one or more velocity representations, the one or more velocity representations including encoded values corresponding to at least one of velocity in the x-direction or velocity in the y-direction.

4. The processor of claim 1 , wherein the calculation of the second data is further based at least in part on third data representing one or more representations corresponding to at least one of an object instance or an object appearance determined using the plurality of outputs. The processor of claim 1 , wherein the fusion of the plurality of rasterized images comprises a fused rasterized image.

6. The processor of claim 5, wherein the one or more objects include at least one of a vehicle, a pedestrian, a cyclist, a motorist, a lane marking, a road boundary marking, a free space boundary, or a waiting line.

7. The processor of claim 1 , wherein: A first output of the plurality of outputs corresponds to a first field of view; a second output of the plurality of outputs corresponds to a second field of view different from the first field of view; And the fusion of the plurality of rasterized images corresponds to both the first field of view and the second field of view.

8. The processor of claim 7, wherein the first field of view and the second field of view at least partially overlap.

9. The processor of claim 1 , wherein the first data further represents one or more additional outputs generated using a LiDAR sensor, a RADAR sensor, or an ultrasonic sensor, and the one or more additional outputs are generated using another DNN or without using another DNN.

10. The processor of claim 1 , wherein: A first output of the plurality of outputs comprises a first representation of an object; a second output of the plurality of outputs comprises a second representation of the object; and a fusion of the plurality of rasterized images comprises a fused representation of the object.

11. The processor of claim 1 , wherein the processor is included in at least one of: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for performing deep learning operations; a system implemented using an edge device; a system implemented using a robot; a system incorporating one or more virtual machines (VMs); a system implemented at least in part in a data center; or a system implemented at least in part using cloud computing resources.

12. A system comprising: one or more processing units; and one or more memory units storing instructions that, when executed by the one or more processing units, cause the one or more processing units to perform operations including: receiving first data representing at least a first rasterized image, the first rasterized image generated using a first deep neural network (DNN) and based at least in part on first sensor data generated using a first sensor, the first rasterized image including at least a first object; receiving second data representing at least a second rasterized image, the second rasterized image generated using a second deep neural network (DNN) and based at least in part on second sensor data generated using a second sensor, the second rasterized image including at least a second object; calculating third data representing a fused rasterized image based at least in part on the first data and the second data using a fused DNN, the fused rasterized image including both the first object and the second object; and performing one or more operations using an autonomous machine based at least in part on the third data.

13. The system of claim 12, wherein the first sensor and the second sensor comprise one of an image sensor, a LiDAR sensor, a RADAR sensor, or an ultrasonic sensor.

14. The system of claim 12, wherein the first sensor and the second sensor include at least partially overlapping fields of view, the first rasterized image includes a first representation of a third object, the second rasterized image includes a second representation of the third object, and the fused rasterized image includes a fused representation of the third object.

15. The system of claim 12, wherein the operations further comprise: receiving fourth data representing at least one probability distribution function corresponding to at least one pixel of at least one of the first rasterized image or the second rasterized image, the at least one pixel corresponding to at least one of the first object or the second object, and the at least one probability distribution function corresponding to one or more potential locations of a detected object, wherein calculating the third data is further based at least in part on the fourth data.

16. The system of claim 12, wherein the operations further comprise: Fourth data representing one or more velocity representations including encoded values corresponding to at least one of velocity in the x-direction or velocity in the y-direction is received, wherein calculating the third data is further based at least in part on the fourth data.

17. The system of claim 12, wherein the system is included in at least one of: a control system for an autonomous or semi-autonomous machine; a perception system for an autonomous or semi-autonomous machine; a system for performing simulation operations; a system for performing deep learning operations; a system implemented using an edge device; a system implemented using a robot; a system incorporating one or more virtual machines (VMs); a system implemented at least in part in a data center; or a system implemented at least in part using cloud computing resources.

18. A method comprising: Receiving first data representing at least a first rasterized image, the first rasterized image generated at least in part based on first sensor data, the first sensor data generated using a first sensor of a first type, the first rasterized image including at least a first object; receiving second data representing at least a second rasterized image, the second rasterized image generated at least in part based on second sensor data, the second sensor data generated using a second sensor of a second type different from the first type, the second rasterized image including at least a second object; calculating third data representing a fused rasterized image based at least in part on the first data and the second data using a fused deep neural network (DNN), the fused rasterized image including both the first object and the second object; and performing one or more operations based at least in part on the third data using an autonomous machine.

19. The method of claim 18, wherein the first type and the second type comprise one of an image sensor, a LiDAR sensor, a RADAR sensor, or an ultrasonic sensor.

20. The method of claim 18, wherein the first type comprises an image sensor and the first rasterized image is generated using a deep neural network (DNN), and the second type comprises one of a LiDAR sensor, a RADAR sensor, or an ultrasonic sensor, and the second rasterized image is generated without using a DNN.

Images