Intersection Posture Detection in Autonomous Machine Applications

Through real-time vehicle perception data and machine learning algorithms, the intersection structure and attitude are automatically detected and navigation paths are generated, which solves the problem of navigation path generation in complex intersection environments in the existing technology, and realizes the efficient driving and system of autonomous vehicles in urban environments. Scalability.

CN113785302BActive Publication Date: 2024-11-26NVIDIA CORP
View PDF 0 Cites 0 Cited by

Patent Information

Application Number
CN202080031631.3
Authority / Receiving Office
CN · China
Patent Type
Patents(China)
Current Assignee / Owner
Priority Date
2019-04-26
Filing Date
2020-04-14
Publication Date
2024-11-26
Estimated Expiration
2040-04-14

AI Technical Summary

Technical Problem

It is difficult for existing autonomous driving systems and ADAS to accurately detect intersection postures and generate navigation paths in complex intersection environments, especially in urban environments, which require a large amount of human markings and map data, resulting in high complexity and no scalability.

Method used

Real-time or near-real-time vehicle perception data is used, and machine learning algorithms such as deep neural networks automatically detect the intersection structure and attitude, generate paths for navigation, and use key points such as lanes, pedestrian crossings, intersection entry or departure lines for path planning.

Benefits of technology

It realizes more free and efficient autonomous vehicle driving in urban environments, reduces dependence on map data, improves the scalability and real-time nature of the system, and reduces the need for human markings.

✦ Generated by Eureka AI based on patent content.

Smart Images

  • Figure CN113785302B_ABST
    Figure CN113785302B_ABST
Patent Text Reader

Abstract

In various examples, scene perception from sensors of a vehicle can be utilized to generate potential paths for the vehicle in order to navigate an intersection in real time or near real time. For example, a deep neural network (DNN) can be trained to compute various outputs such as heat maps corresponding to key points associated with the intersection, vector fields corresponding to directionality, heading, and offset relative to lanes, intensity maps corresponding to the width of lanes, and / or classifications corresponding to line segments of the intersection. The outputs can be decoded and / or otherwise post-processed to reconstruct the intersection (or key points corresponding thereto) and to determine a suggested or potential path for navigating the vehicle through the intersection.
Need to check novelty before this filing date? Find Prior Art

Description

Background Art

[0001] Autonomous driving systems and advanced driver assistance systems (ADAS) can utilize sensors (such as cameras, LIDAR sensors, RADAR sensors, etc.) to perform different tasks - such as lane keeping, lane changing, lane assignment, camera calibration, turning, path planning and / or positioning. For example, in order for autonomous systems and ADAS systems to operate independently and efficiently, an understanding of the vehicle's surroundings must be achieved. This understanding can include information about the location of objects, obstacles, lanes, intersections, etc., and this information can be used by the vehicle when making path planning or control decisions - such as what path or trajectory to follow.

[0002] For example, information about the location and layout of intersections in the environment of an autonomous or semi-autonomous vehicle is important when making path planning, obstacle avoidance, and / or control decisions (such as where to stop, what path to use to safely cross an intersection, where other vehicles or pedestrians may be located, etc.). This is particularly important when the vehicle is operating in an urban and / or suburban driving environment, where intersection understanding and path planning become critical due to the increased number of variables relative to a highway driving environment. For example, in the case of a vehicle executing a left turn through an intersection in a two-way, multi-lane driving environment, determining the position and directionality relative to other lanes and determining the location of crosswalks or bicycle lanes becomes critical for safe and effective autonomous and / or semi-autonomous driving.

[0003] In conventional systems, intersections may be interpolated from a pre-stored high-definition (HD), three-dimensional (3D) map of the vehicle. For example, the structure and orientation of the intersection and surrounding area may be gleaned from an HD map. However, in addition to marking each intersection pose feature that may be associated with each intersection, such a map-based solution also requires accurate and detailed human labeling of each intersection to be navigated. Therefore, the more complex the intersection, the more annotations are required, thereby increasing the complexity of accurately labeling the intersection and reducing the scalability of intersection structure detection using a map-based approach. For example, the process may be logistically more complex when manual labeling of larger or denser geographic areas (e.g., cities, states, countries) is required so that vehicles can independently and efficiently drive in different areas. Summary of the invention

[0004] Embodiments of the present disclosure relate to intersection posture detection in autonomous machine applications. Systems and methods are disclosed that utilize outputs from various sensors of a vehicle to automatically detect and generate a path to navigate an intersection in real-time or near real-time.

[0005] Compared with conventional systems (such as those described above), current systems can use the real-time perception of vehicles to detect intersection gestures and generate paths for navigating intersections. Key points (e.g., center points and / or endpoints) of line segments corresponding to features of intersections (such as lanes, crosswalks, intersection entry or exit lines, bicycle lanes, etc.) can be used to generate potential paths for vehicle navigation intersections. For example, one or more machine learning algorithms (such as deep neural networks (DNNs)) can be trained to calculate information corresponding to intersections (such as key points, heading directions, widths of lanes, number of lanes, etc.), and this information can be used to connect central key points (e.g., key points corresponding to the center of line segments) together to generate paths and / or tracks for vehicles to effectively and accurately navigate intersections. In this way, semantic information associated with predicted key points (such as directionality, heading, width and / or classification information corresponding to the section of the intersection) can be calculated and utilized to obtain an understanding of the intersection gesture. For example, the output of the DNN can be used to directly or indirectly (e.g., via decoding) determine: the location of each lane, bike lane, crosswalk, etc.; the number of lanes associated with the intersection; the geometry of the lanes, bike lanes, crosswalks, etc.; the direction of travel (or heading direction) corresponding to each lane; and / or other intersection structure information.

[0006] As a result of using situational awareness to generate an understanding of each intersection, the process of generating a path for navigating the intersection may be relatively less time consuming, less computationally intensive, and more scalable because the system can learn to diagnose each intersection in real time or near real time, without requiring prior experience or knowledge of the intersection. As a result, autonomous vehicles may be able to more freely travel through cities, urban environments, or other locations where HD maps may not be readily available or fully accurate. BRIEF DESCRIPTION OF THE DRAWINGS

[0007] The present system and method for intersection posture detection in autonomous machine applications are described in detail below with reference to the accompanying drawings, wherein:

[0008] Figure 1 is a data flow diagram illustrating an example process for training a neural network to detect intersection pose information for navigating an intersection using output from sensors of a vehicle according to some embodiments of the present disclosure;

[0009] Figure 2A-2B is an illustration of example annotations of images for generating ground truth data for training a neural network to detect intersection structure and pose according to some embodiments of the present disclosure;

[0010] Figure 3A is an illustration of example ground truth data generated from annotations for encoding key points according to some embodiments of the present disclosure;

[0011] Figure 3B is an illustration of an example calculation for generating an offset vector for ground truth data corresponding to a keypoint according to some embodiments of the present disclosure;

[0012] Figure 3C is an illustration of a portion of an example vector field for encoding heading and / or directionality relative to a lane or line segment according to some embodiments of the present disclosure;

[0013] Figure 4 is a flow chart illustrating an example method for training a neural network to detect intersection structures according to some embodiments of the present disclosure;

[0014] Figure 5 is a data flow diagram illustrating an example process for detecting and generating a path to navigate an intersection in real time or near real time using output from sensors of a vehicle according to some embodiments of the present disclosure;

[0015] Fig. 6A is a visualization of example intersection structures and poses generated using a neural network according to some embodiments of the present disclosure;

[0016] Figure 6B is an illustration of an example path generated by a path generator based on intersection structure and pose as computed using a neural network according to some embodiments of the present disclosure;

[0017] Figure 7 is a flow chart illustrating an example method for calculating intersection structure and pose and a path through an intersection according to some embodiments of the present disclosure;

[0018] Fig. 8A is an illustration of an example autonomous vehicle according to some embodiments of the present disclosure;

[0019] Figure 8B According to some embodiments of the present disclosure Fig. 8A Examples of autonomous vehicle camera positions and fields of view;

[0020] Figure 8C According to some embodiments of the present disclosure Fig. 8A a block diagram of an example system architecture for an example autonomous vehicle;

[0021] Fig.8D According to some embodiments of the present disclosure, a method for Fig. 8A A system diagram of an example of communication between autonomous vehicles; and

[0022] Fig. 9 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0023] Systems and methods are disclosed that relate to intersection posture detection in autonomous machine applications. Although the present disclosure may be related to an example autonomous vehicle 800 (alternatively referred to herein as “vehicle 800” or “autonomous vehicle 800”), the example Figures 8A-8D The systems and methods described herein may be described with reference to, but are not intended to be limiting. For example, the systems and methods described herein may be used by, but are not limited to, non-autonomous vehicles, semi-autonomous vehicles (e.g., in one or more adaptive driver assistance systems (ADAS)), robots, warehouse vehicles, off-road vehicles, spacecraft, boats, shuttles, emergency response vehicles, motorcycles, electric bicycles or motorized bicycles, airplanes, construction vehicles, underwater aircraft, drones, and / or other vehicle types. Furthermore, while the present disclosure may describe intersection structures for vehicle applications, this is not intended to be limiting, and the systems and methods described herein may be used for augmented reality, virtual reality, robotics, security and surveillance, autonomous or semi-autonomous machine applications, and / or any other technology space where detection of intersections or other environmental structures and / or gestures may be used.

[0024] As described herein, compared to conventional methods of intersection structure and posture detection using high-definition (HD) maps, current systems and methods provide techniques for detecting and generating paths for navigating intersections in real time or near real time using outputs from sensors (e.g., cameras, RADAR sensors, LIDAR sensors, etc.) of vehicles. As such, for each intersection, the vehicle's real-time perception can be used to detect intersection postures and generate paths for navigating intersections. To this end, embodiments of the present disclosure can utilize key points (e.g., midpoints and / or endpoints) of line segments corresponding to lanes, crosswalks, pedestrian intersections, etc. to generate one or more potential paths for vehicle navigation intersections. Computer vision and / or machine learning algorithms can be trained to detect key points of intersections, and (center) key points can be connected together (using one or more filters) to generate paths and / or trajectories for vehicles to effectively and accurately navigate intersections. To generate a path, directionality, heading, width, and / or semantic information corresponding to segments of an intersection may be computed and utilized in order to gain an understanding of the intersection's pose—e.g., the location and number of lanes, the direction of travel of each lane, etc. As such, by using perception to generate an understanding of each intersection, the process of generating a path for navigating intersections may be relatively less time consuming, less computationally intensive, and more scalable because the system may learn to diagnose each intersection in real time or near real time without requiring prior experience or knowledge of the intersection. In some embodiments, these perception-based methods may be performed in conjunction with map-based methods to provide redundancy and to further validate the results of the perception-based methods.

[0025] In a deployment, sensors located or otherwise arranged on an autonomous or semi-autonomous vehicle (e.g., cameras, RADAR sensors, LIDAR sensors, etc.) may be used to receive and / or generate sensor data (e.g., images, videos, etc.). The sensor data may be applied to a neural network (e.g., a deep neural network (DNN) such as a convolutional neural network (CNN)) that is trained to identify regions of interest associated with road signs, road boundaries, intersections, etc. (e.g., raised pavement markings, rumble strips, colored lane dividers, sidewalks, crosswalks, forks, etc.) represented by the sensor data, as well as semantic and / or directional information associated therewith. More specifically, the neural network can be designed to calculate key points corresponding to sections of the intersection (e.g., corresponding to lanes, bike lanes, etc., and / or corresponding to features therein, such as crosswalks, intersection entry points, intersection exit points, etc.), and generate outputs that identify, for each key point, the width of the lane corresponding to the key point, the directionality of the lane, the heading corresponding to the lane, the semantic information corresponding to the key point (e.g., crosswalk, crosswalk entrance, crosswalk exit, intersection entrance, intersection exit, etc., or a combination thereof), and / or other information. In some examples, the calculated key points can be represented by pixels represented by sensor data, and the centers and / or endpoints of the intersection features are located at these pixels, such as pedestrian (e.g., crosswalk) entrance lines, intersection entrance lines, intersection exit lines, and pedestrian exit lines. In this way, the key points and related information can be used to determine the structure and posture of the intersection, where the structure may include representations of lanes, lane types, key points, crosswalks, directionality, headings, and / or other information corresponding to the intersection.

[0026] The DNN can be trained to predict any number of different information corresponding to the structure, orientation, and posture of the intersection (e.g., via any number of lanes). For example, a lane may represent a key point heat map, an offset vector, a direction vector, a heading vector, the number of lanes and lane types, the width of the lane, and / or other lanes. During training, the DNN may be trained with images or other sensor data representations labeled or annotated with line segments representing lanes, crosswalks, entrance lines, exit lines, bike lanes, etc. (and may further include semantic information corresponding thereto). The ground truth encoder can then use the labeled line segments and semantic information to generate key point heat maps, offset vector maps, direction vector maps, heading vector maps, lane width intensity maps, lane counts, lane classifications, and / or other information corresponding to the intersection determined from the annotations. In some examples, key points may also include corners or endpoints of line segments corresponding to lanes, which may be inferred from center points, line direction vectors, and / or width information, or may be directly calculated using one or more vector fields corresponding to endpoint key points. Therefore, these keypoints can be used to encode the intersection structure with limited labeling required, as the information can be determined using the line segment annotations and the associated semantic information.

[0027] As a result of different intersections including different numbers of associated keypoints, many different two-dimensional (2D) encoding maps can be used to encode the ground truth data so that each training image has the same ground truth dimensions (e.g., rather than creating a maximum number of keypoint predictions and marking non-existent keypoints as null). As an example, the heatmaps described herein can be used to encode keypoint locations, where each keypoint can be represented by a non-normalized Gaussian distribution with a mean parameter equal to the keypoint location. However, unlike standard Gaussian heatmaps, in some embodiments, directional and adaptive variance Gaussian heatmaps can be encoded - where the variance is set according to the ground truth lane width in image space, and the covariance is calculated from the ground truth line segment directions. For example, a heatmap may include a non-normalized normal distribution with a fixed standard deviation and / or an adaptive covariance.

[0028] In addition to heatmaps of keypoint locations, two-dimensional (2D) vector fields can be used to encode heading direction and line direction. 2D line direction vector fields and 2D heading direction vector fields can be encoded by assigning a direction vector to a center keypoint and then assigning the same direction vector to each pixel within a defined radius of the center keypoint's location. The lane width can be encoded by assigning an intensity value equal to the lane width normalized by the image width, and the same intensity value can be assigned to other pixels within the defined radius. In some examples, the endpoint keypoints can be encoded using an offset vector field so that the DNN learns to calculate not only the location of the center keypoint, but also the location of the endpoint keypoints. Using this information, the width of the lane can be determined directly from the endpoint keypoints (e.g., the distance from one endpoint to the other), the directionality of a line segment extending across the lane can be determined from the endpoint keypoints (e.g., the direction can be defined by a line from one endpoint to the other), and / or the heading direction can be determined from the endpoint keypoints (e.g., taking a normal from the directionality to define an angle, and using semantic information to determine the heading).

[0029] Furthermore, because the 2D encoding pass may not have the same spatial resolution as the input image (e.g., because the image may be downsampled during processing by the DNN), a smaller encoding pass may be used. To compensate for information loss during downsampling, the offset vector field may be encoded by assigning a vector pointing to the closest ground truth keypoint pixel location (e.g., to the closest center keypoint, left end (or left edge) keypoint, and / or right end (or right edge) keypoint) for each pixel in the offset vector field. This may allow the DNN to preserve the original positions of the keypoints in the input image at the input spatial resolution.

[0030] Once the DNN is trained, the DNN can be regressed on the output in the form of a Gaussian heat map, a direction vector field, a heading vector field, a lane width density map, semantic information, an offset vector field, and / or other outputs corresponding to the encoded ground truth channel that the network is trained to predict, as described herein. As a result of the DNN output (which includes heat maps, vector fields, density maps, etc.), the key points in the output can be represented by a probability distribution that needs to be post-processed into a spatial distribution of 2D coordinates, and the 2D coordinates represent the location of the key points (e.g., the center key point, the left end key point, and / or the right end key point). In some embodiments, a peak finding algorithm can be applied to determine the point with the highest probability within a sliding window of a predetermined size. The final result may include the pixel location corresponding to the peak of the probability distribution achieved by non-maximum suppression. The 2D location of the surviving peak probability point can be output (e.g., using a post-processing module) along with the corresponding semantic information as determined by the DNN-based classification. In some embodiments, the points can be further filtered to remove overlap using non-maximum suppression or another filtering algorithm.

[0031] The final output of the DNN and post-processing modules may include key points in 2D image space with corresponding semantic classifications. These 2D image space points can be used to determine the three-dimensional (3D) world space coordinates of the key points to help the vehicle navigate through the world space environment. However, in some embodiments, the DNN can be trained on 3D data to predict positions in 3D world space.

[0032] Once the key points are output by the DNN and the 3D positions are determined, any number of additional post-processing steps may be performed in order to finally "connect the points (or key points)" and generate a path for navigating the intersection. For example, the central key points may be connected to generate a polyline representing a potential path for crossing the intersection. The final path may be assigned a path type, which may be determined relative to the position of the vehicle, the position of the key points, and / or the heading vector (e.g., angle). Potential path types include left turns, right turns, lane switching, and / or continuing in a lane. Curve fitting may also be implemented to determine the final shape of the natural curve that most accurately reflects the potential path. Curve fitting may be performed using polyline fitting, polynomial fitting, convolution fitting, and / or other types of curve fitting algorithms. The shape of the potential path may be determined based on the position and heading vector (e.g., angle) associated with the key points to be connected. In an embodiment, the shape of the potential path may be aligned with a tangent to the heading vector. The curve fitting process may be repeated for all key points that may potentially be connected to each other to generate all possible paths that the vehicle may take to navigate the intersection. In some examples, infeasible paths may be removed from consideration based on traffic regulations and physical limitations associated with such paths.

[0033] In some embodiments, a matching algorithm can be used to connect key points and generate a potential path for a vehicle navigation intersection. In these examples, the matching score of each pair of key points can be determined based on the shape of the fitting curve between the position of the key points, the heading vector, and the key point pair. In some examples, a linear matching algorithm can be used, such as the Hungarian sub-matching algorithm. In other examples, a nonlinear matching algorithm (such as a spectral matching algorithm) can be used to connect a pair of key points.

[0034] In either example, once a path through an intersection is determined, the information can be used to perform one or more operations by the vehicle. For example, a world model manager can update a world model to assist in navigating the intersection, a path planning layer of an autonomous driving software stack can use the intersection information to determine a path through the intersection (e.g., along one of the determined potential paths), and / or a control component can determine controls for the vehicle to navigate the intersection based on the determined path.

[0035] See also Figure 1 , Figure 1 is an example data flow diagram illustrating an example process 100 for training a neural network to detect intersection structure and posture using outputs from one or more sensors of a vehicle in accordance with some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are set forth merely as examples. At a high level, process 100 may include one or more machine learning models 104 that receive one or more inputs such as sensor data 102 and generate one or more outputs such as one or more outputs 106. In some examples, when used for training, sensor data 102 may be referred to as training data. Although sensor data 102 is primarily discussed with respect to image data representing images, this is not intended to be limiting, and sensor data 102 may include other types of sensor data for intersection posture detection, such as LIDAR data, SONAR data, RADAR data, etc., for example, as generated by one or more sensors of vehicle 800 ( FIG. 8A to FIG. 8D ).

[0036] Process 100 may include generating and / or receiving sensor data 102 from one or more sensors. As a non-limiting example, sensor data 102 may be received from a vehicle (e.g., FIG. 8A to FIG. 8C The sensor data 102 may be received by one or more sensors of the vehicle 800 and the vehicles described herein. The sensor data 102 may be used by the vehicle and within the process 100 to detect and generate a path to navigate the intersection in real time or near real time. The sensor data 102 may include, but is not limited to, sensor data 102 from any sensor of the vehicle, such as and referring to Figures 8A-8C, one or more global navigation satellite system (GNSS) sensors 858 (e.g., one or more global positioning system sensors), one or more RADAR sensors 860, one or more ultrasonic sensors 862, one or more LIDAR sensors 864, one or more inertial measurement unit (IMU) sensors 866 (e.g., one or more accelerometers, one or more gyroscopes, one or more magnetic compasses, one or more magnetometers, etc.), one or more microphones 876, 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 mid-range cameras 878, one or more speed sensors 844 (e.g., for measuring the speed of the vehicle 800), and / or other sensor types. As another example, the sensor data 102 may include virtual sensor data generated from any number of sensors of a virtual vehicle or other virtual object. In such an example, the virtual sensor may correspond to a virtual vehicle or other virtual object in a simulated environment (e.g., for testing, training, and / or validating neural network performance), and the virtual sensor data may represent sensor data captured by the virtual sensor within the simulated or virtual environment. As such, by using virtual sensor data, one or more machine learning models 104 described herein may be tested, trained, and / or validated using simulated data in a simulated environment, which may allow for testing of more extreme scenarios outside of a real-world environment where such testing may be less safe.

[0037] In some embodiments, the sensor data 102 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 sensing field of a sensor (e.g., a depth map of a LIDAR sensor, a value map of an ultrasonic sensor, etc.). Where the sensor data 102 includes image data, any type of image data format may be used, such as, but not limited to, compressed images such as Joint Photographic Experts Group (JPEG) or Luma / Chrominance (YUV) formats, compressed images as frames derived from a compressed video format such as H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC), raw images such as derived from Red-Transparent-Blue (RCCB), Red-Transparent-Blue (RCCC), or other types of imaging sensors, and / or other formats. Additionally, in some examples, sensor data 102 may be used within process 100 without any preprocessing (e.g., in a raw or captured format), while in other examples, sensor data 102 may undergo preprocessing (e.g., noise balancing, demosaicing, scaling, cropping, enhancement, white balancing, tone curve adjustment, etc., such as using a sensor data preprocessor (not shown)). As used herein, sensor data 102 may refer to unprocessed sensor data, preprocessed sensor data, or a combination thereof.

[0038] Sensor data 102 for training may include original images (e.g., as captured by one or more image sensors), downsampled images, upsampled images, cropped images or region of interest (ROI) images, otherwise enhanced images, and / or combinations thereof. One or more machine learning models 104 may be trained using images (and / or other sensor data 102) and corresponding ground truth data 122. The ground truth data 122 may include annotations, labels, masks, maps, and / or the like. For example, in some embodiments, the ground truth data 122 may include one or more intensity maps 128, one or more heat maps 124, and / or one or more vector fields 126, such as direction vector fields 126A, heading vector fields 126B, and / or offset vector fields 126C. The encoder 120 may encode the ground truth data 122 using annotations 118. In some embodiments, the annotations 118 may include one or more lane labels 118A and one or more classifications 118B of one or more lane labels 118A.

[0039] Referring to the annotations 118, in some examples, the annotations 118 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 the annotations 118 and / or other ground truth data 122, and / or may be hand-drawn. In any example, the annotations 118 and / or the ground truth data 122 may 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), manually annotated (e.g., a labeler or annotation expert who defines the location of the labels), and / or a combination thereof (e.g., a person identifies the center or origin and dimensions of the lane, and a machine generates the lane).

[0040] Lane (or line) labels 118A may include annotations or other label types corresponding to features or areas of interest corresponding to the intersection. In some examples, the intersection structure may be defined as a set of line segments corresponding to lanes, sidewalks, entrance lines, exit lines, bicycle lanes, etc. in the sensor data 102. The line segments may be generated as polylines, where the center of each polyline is defined as the center of the corresponding line segment. One or more classifications 118B may be generated for each image in the image (or other data representation) and / or for each one or more line segments and centers in the image represented by the sensor data 102 used to train one or more machine learning models 104. The number of one or more classifications 118B may correspond to the number and / or type of features that one or more machine learning models 104 are trained to predict, or to the number and / or type of lanes in the corresponding image. Depending on the embodiment, one or more classifications 118B may correspond to classifications or labels corresponding to feature types, such as, but not limited to, crosswalks, crosswalk entrances, crosswalk exits, intersection entrances, intersection exits, and / or bicycle lanes.

[0041] In some examples, the intersection structure may be determined based on the annotations 118. In these examples, a set of key points may be determined from the lane labels 118A, where each key point corresponds to the center (or left edge or right edge, etc.) of the corresponding lane. Although the key points are primarily described relative to the center points of the lane segments, this is not intended to be limiting, and in some examples, the corners or endpoints of each lane may also be determined as key points for the corresponding instance of the sensor data 102. For example, the corners or endpoints of each lane may be inferred from the center key point and directionality of the lane or from the lane labels 118 themselves. In addition, the number of lanes and the heading, directionality, width, and / or other geometric structures corresponding to each lane may be determined from the annotations 118 (e.g., from one or more lane labels 118A and one or more classifications 118B). Therefore, even though the annotations 118 may not directly indicate certain intersection structure or posture information (such as the heading direction of the lane, lane width, and / or directionality), the annotations 118 may be analyzed to determine this information.

[0042] For example, where a first lane label 118A extends along the width of the lane and includes a classification of "sidewalk entrance", and a second lane label 118A extends along the width of the same lane and includes a classification of "sidewalk exit_intersection entrance", this information can be used to determine the heading direction of the lane (e.g., from the first lane label to the second lane label) (e.g., the vehicle will be traveling towards the second lane label in a direction across the first lane label). In addition, lane label 118A can indicate the directionality (e.g., angle) of the lane, and based on this information, the heading direction (e.g., angle) can be determined - such as by calculating the normal of lane label 118.

[0043] See now Figure 2A-2B , Figure 2A-2B Example annotations applied to sensor data for use in ground truth generation for training machine learning models to detect intersection structure and pose according to some embodiments of the present disclosure are shown. Figure 2A Example labeling of the illustrated image 200A (eg, corresponding to the annotation 118) that may be used to identify the image 200A according to the embodiment of the present invention. Figure 1The training process 100 generates ground truth data. Lanes in the image can be annotated with lane labels 118A (e.g., lanes 204, 206, 210, and 212) and corresponding classifications (e.g., pedestrian entrance, intersection exit, intersection entrance, pedestrian exit, no lane). For example, lane 204 can be marked with a line segment and classified as one or more of an intersection entrance and a pedestrian entrance. Similarly, lane 206, lane 208, and lane 210 can be marked with a line segment, wherein lane 206 can be classified as one or more of an intersection entrance and a pedestrian exit, lane 208 can be classified as one or more of an intersection entrance and a pedestrian exit, and lane 212 can be classified as no lane.

[0044] Further, the labels for lanes 204, 206, and 208 may be further annotated with corresponding heading directions, as indicated by arrows 202A-202V. The heading direction may indicate the direction of traffic associated with a particular lane. In some examples, the heading direction may be associated with the center (or key) point of its corresponding lane label. For example, heading direction 202S may be associated with the center point of lane 204. Figure 2A Different classification labels may be represented by different line types (e.g., solid line, dashed line, etc.) to represent different classifications. However, this is not intended to be limiting, and any visualization of lane labels and their classifications may include different shapes, patterns, fills, colors, symbols, and / or other identifiers for illustrating differences in classification labels of features (e.g., lanes) in the image.

[0045] See now Figure 2B , Figure 2B Another example of annotations applied to sensor data to train a machine learning model to detect intersection structure and posture according to some embodiments of the present invention is shown. Here, lanes 222A and 222B in image 200B may be annotated with line segments. The line segment corresponding to lane 222B may be annotated to extend across the vehicle. This can help train one or more machine learning models 104 to predict the location of key points even in cases where the actual location may be obscured. As a result, the presence of vehicles or other objects in the road may not detract from the system's ability to generate suggested paths through the intersection. For the same classification, the annotations may have similar visual representations. As shown, lanes 222A and 222B may be classified as intersection entry lines stop lines. In this way, features of similar classifications of an image may be annotated in a similar manner. Furthermore, it should be noted that one or more classifications 118B may be compound nouns. In Figure 2BIn the figure, different classification labels may be represented by solid lines, dashed lines, etc. to represent different classifications. Further, different classification labels may be nouns and / or compound nouns. This is not intended to be limiting, and any naming convention for classification may be used to illustrate differences in classification labels for features (e.g., lanes) in an image.

[0046] See again Figure 1 , the encoder 120 may be configured to use the annotations 118 to encode ground truth information corresponding to the structure and posture of the intersection. For example, as described herein, even though the annotations may be limited to lane labels 118A and classifications 118B, information such as key points, number of lanes, heading direction, directionality, width, and / or other structure and posture information may be determined from the annotations 118. Once this information is determined, it may be encoded by the encoder 120 to produce ground truth data 122. In addition, because the spatial resolution of the sensor data 102 applied to one or more machine learning models 104 may be different from (e.g., greater than, less than, etc.) the output spatial resolution corresponding to the prediction of the one or more machine learning models 104, the ground truth information may be encoded into a variety of formats.

[0047] For example, key points (e.g., center points, inflection points, endpoints, etc.) can be encoded in the form of a heat map 124 so that each training image of the sensor data 102 has the same ground truth dimensions (e.g., rather than creating a certain maximum number of key point predictions and marking non-existent points as null). In this way, images of different intersections with different numbers of associated key points can be encoded to include the same ground truth dimensions, for example, by representing the key points in the form of a heat map 124. For non-limiting examples, each key point can be represented by a non-normalized Gaussian distribution with a mean parameter equal to the key point position. However, unlike a standard Gaussian heat map, in some embodiments, a directional and adaptive variance Gaussian heat map (e.g., a Gaussian heat map with an adaptive directional radius) can be encoded to produce a heat map 124. In such an example, one or more heat maps 124 can be generated using a variance set based on the ground truth lane width in image space and the covariance calculated from the ground truth line segment direction. For example, a heat map can include a non-normalized normal distribution with a fixed standard deviation and / or an adaptive covariance. In some examples, heat map 124 may include multiple heat maps, where each heat map corresponds to a different type of classification 118B of key points determined from annotations 118 .

[0048] As an example, refer to Figure 3A , Figure 3AA directional and adaptive covariance heatmap 300C is shown generated using an annotated image 300A. The annotated image 300A includes key points 302. In some examples, the key points 302 may include center points of image features (e.g., line segments corresponding to lanes). In some other examples, the key points 302 may also include endpoints of line segments. The directional and adaptive covariance heatmap 300C and the standard Gaussian heatmap are two non-limiting examples of heatmaps 124 that may be used to encode key points corresponding to line segments. For example, the standard Gaussian heatmap 300B may encode key point locations using a fixed variance and no correlation between 2D directions. The directional and adaptive variance heatmap 300B may encode key points, where the variance of each classification is set according to the lane width in the image space corresponding to the lane, and the covariance is determined from the line direction of the corresponding lane (e.g., left and right directions pointing to the endpoints). Each key point in the heatmap may be encoded as represented by a non-normalized Gaussian distribution with a mean parameter equal to the key point location. As Figure 3A As shown, each semantic category of a key point is associated with a different pattern within the heat map. For example, a key point 304A in one semantic category has a different pattern than a key point 310A in another semantic category. In embodiments, a different heat map may be generated for each semantic category of a key point, while in other embodiments, a single heat map may represent each semantic category with some variation within the heat map corresponding to a particular semantic category.

[0049] In some embodiments, a left end keypoint and a right end keypoint, corresponding to the left edge of the lane and the right edge of the lane, respectively, may be encoded in a vector field, e.g., a direction vector field 126A. For example, instead of or in addition to encoding the directionality of a line segment in the direction vector field 126A (as described herein), the direction vector field 126A may represent a vector to a left end keypoint, a center keypoint, and / or a right end keypoint. As such, one or more machine learning models 104 may be trained to output a direction vector field 112A that may be decoded to determine the location of a left end keypoint, a center keypoint (e.g., in addition to or in lieu of a heat map 108), and / or a right end keypoint. In some examples, the direction vector field 126A may be encoded as a decomposed vector, wherein the vector is decomposed into a unit norm direction vector and a magnitude (e.g., a length).

[0050] As an example, and relative to Figure 3B, to encode the ground truth direction vector field 126A, a pixel (x, y) 338 at the output resolution (e.g., a sparse output, e.g., due to downsampling) may be encoded to include a first vector 326 from the pixel 338 to the left keypoint 320, a second vector 330 from the pixel 338 to the right keypoint 324, and / or a third vector 328 from the pixel 338 to the center keypoint 322. The left keypoint 320, the right keypoint 324, and the center keypoint 322 may represent the positions of these keypoints at the output resolution of the machine learning model 104, as determined from their corresponding positions at the input resolution of the sensor data 102. For example, because the output resolution may be downsampled relative to the input resolution, there may not be a 1:1 pixel correspondence at the output resolution. As such, vectors 326, 328, and 330 may encode the locations of keypoints 320, 324, and 322 at the output resolution (and / or the ground truth resolution) so that predictions of the machine learning model 104 (e.g., the computed direction vector map 112A) may be projected back to the input resolution. Keypoints 320, 322, and 324 may correspond to polylines 334 annotated with respect to the ground truth sensor data at the input resolution.

[0051] In the example where the left end key point and the right end key point are encoded in the direction vector field 126A, the heading direction and / or width of the lane can be directly determined from the reconstructed polyline. For example, the direction vector field 112A can be decoded to determine the left end key point and the right end key point corresponding to the line segment, and each line segment can be reconstructed by forming a line between the left end key point and the right end key point (and in an embodiment, the center key point decoded from the heat map 108). In this way, the width of the lane corresponding to the line segment can be determined from the length or magnitude of the reconstructed line segment (and / or from the distance between the left end key point and the right end key point). In addition, the normal of the direction of the line segment can be used to determine the heading angle corresponding to the lane, and the heading direction can be determined using semantic or classification information (for example, in the case where the line segment corresponds to the entrance of the crosswalk and the next line segment after the line segment corresponds to the exit of the crosswalk and the entrance of the intersection, the heading direction can be determined to be from the line segment toward the next line segment, and the heading is to enter the intersection).

[0052] In some embodiments, the heading direction and / or directionality of the line segment may be encoded into the direction vector field 126, where the vector corresponds to the actual directionality of the line segment and / or the heading direction of the lane. For example, instead of encoding the vector to the left end keypoint, the center keypoint, and / or the right end keypoint, or in addition to encoding the vector to the left end keypoint, the center keypoint, and / or the right end keypoint, the vector may represent the angle of the line segment and / or the angle of the heading. As such, the line segment may be decoded by determining the location of the center keypoint from the heat map 108, determining the width of the lane from the intensity map 114, and using the directionality decoded from the direction vector field 112A to generate a line segment extending from the center keypoint to the right and left sides with a total length equal to the width of the lane. With respect to ground truth encoding, for example, once the directionality (e.g., the angle or other geometry of the lane) is determined from the annotation 118, a direction vector may be determined and attributed to the line segment, such as to a keypoint representing the line segment (e.g., the center keypoint). Direction vector field 126A may therefore represent direction vectors of surrounding pixels or points within a radius attributable to the keypoint and / or several additional direction vectors. Figure 3C , a direction vector may be applied to one or more pixels 340 corresponding to one or more keypoints (or the closest corresponding pixels 340, where the output resolution of the one or more machine learning models 104 is different than the input resolution of the sensor data 102 and / or ground truth data), and the same direction vector may also be applied to pixels within a predetermined radius 342 of the keypoint pixel. Each of these direction vectors for each of the keypoint pixels may then be encoded to generate a direction vector field 126A.

[0053] Similarly, once the heading is determined (e.g., from the directionality (as its normal), or otherwise), a heading vector may be determined and attributed to the line segment, such as to a keypoint representing the line segment. For example, once the heading is determined (e.g., an angle corresponding to the direction of travel of the vehicle along a lane), a heading vector may be determined and attributed to the line segment, such as to a keypoint representing the line segment. The heading vector field 126B may thus represent the heading vectors attributed to surrounding pixels or points within a radius of one or more keypoints and / or several additional heading vectors. For example, and relative to Figure 3C , a heading vector may be applied to one or more pixels 340 corresponding to one or more keypoints (or the closest corresponding pixel 340 where the output resolution of one or more machine learning models 104 is different from the input resolution of sensor data 102 and / or ground truth data), and the same heading vector may also be applied to pixels within a predetermined radius 342 of the keypoint pixel. Each of these heading vectors for each of the keypoint pixels may then be encoded to generate a heading vector field 126A.

[0054] In some non-limiting embodiments, the offset vector may be determined and encoded to generate an offset vector field 126C. The offset 126 vector field may be encoded by assigning a vector pointing to the closest ground truth keypoint pixel position to each pixel in the offset vector field. In this way, a smaller encoding channel may be used even when the 2D encoding channel has a different spatial resolution than the input image (e.g., a downsampled image). This allows one or more machine learning models 104 to train and predict intersection structure and posture in a computationally less expensive manner because a smaller encoding channel may be used without losing information due to downsampling of the image by one or more machine learning models 104 during processing. In some examples, the ground truth data 122 for multiple features (e.g., lanes) per classification 118B may be directly encoded using simple counts. One or more machine learning models 104 may then be trained to directly predict the number of features per classification 118B.

[0055] In some embodiments, one or more intensity maps 128 may be implemented to encode lane width as determined from one or more lane labels 118A corresponding to segments of one or more lanes. For example, once the lane width is determined, the lane width of the lane segment corresponding to each keypoint may be encoded by assigning to the keypoint an intensity value equal to the lane width (e.g., in image space) normalized by the image width (e.g., also in image space). In some examples, the same intensity value may be assigned to other pixels within a defined radius of the associated keypoint, similar to as described herein with respect to direction vector field 126A and heading vector field 126B.

[0056] Once the ground truth data 122 is generated for each instance of the sensor data 102 (e.g., for each image where the sensor data 102 includes image data), the ground truth data 122 may be used to train one or more machine learning models 104. For example, the one or more machine learning models 104 may generate one or more outputs 106 that may be compared to the ground truth data corresponding to the respective instances of the sensor data 102 using one or more loss functions 130. As such, feedback from the one or more loss functions 130 may be used to update parameters (e.g., weights and biases) of the one or more machine learning models 104 in light of the ground truth data 122 until the one or more machine learning models 104 converge to an acceptable or desired accuracy. Using the process 100, the one or more machine learning models 104 may be trained to accurately predict the one or more outputs 106 (and / or associated classifications) from the sensor data 102 using the one or more loss functions 130 and the ground truth data 122. In some examples, different loss functions 130 may be used to train one or more machine learning models 104 to predict different outputs 106. For example, a first loss function 130 may be used to compare one or more heat maps 108 and 124, and a second loss function 130 may be used to compare intensity map 128 to intensity map 114. As such, in a non-limiting embodiment, one or more of the output channels is trained using a different loss function 130 than another of the output channels.

[0057] One or more machine learning models 104 may use sensor data 102 to compute one or more outputs 106, which may ultimately be applied to a decoder or one or more other post-processing components (at least with respect to the present invention). Figure 5 ), to generate key points, classifications, lane widths, number of lanes, lane headings, lane directionality, and / or other information. Although the present disclosure relates to using a deep neural network (DNN), and in particular a convolutional neural network (CNN), as a machine learning model 104 (e.g., Figure 1 and Figure 5) to describe an example, but this is not intended to be limiting. For example, but not limited to, one or more machine learning models 104 may include any type of machine learning model, such as a machine learning model using linear regression, logistic regression, decision tree, support vector machine (SVM), naive Bayes, k-nearest neighbor (Knn), K-means clustering, random forest, dimensionality reduction algorithm, gradient boosting algorithm, neural network (e.g., autoencoder, convolution, recursion, perceptron, long / short-term memory / LSTM, Hopfield, Boltzmann, deep belief, deconvolution, generative adversarial, liquid machine, etc.), lane detection algorithm, computer vision algorithm, and / or other types of machine learning models.

[0058] As an example, such as in the case where one or more machine learning models 104 include a CNN, one or more machine learning models 104 may include any number of layers. One or more of the layers may include an input layer. The input layer may hold values ​​associated with the sensor data 102 (e.g., before or after post-processing). For example, when the sensor data 102 is an image, the input layer may hold values ​​representing the raw pixel values ​​of the image as a volume (e.g., width, height, and color channels (e.g., RGB), such as 32x32x3).

[0059] One or more layers may include a convolutional layer. A convolutional layer may compute the output of neurons connected to local regions in the 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 a convolutional layer may be another volume where one dimension is based on the number of filters applied (e.g., the width, height, and number of filters, such as 32x32x12 if 12 is the number of filters).

[0060] One or more layers may include a rectified linear unit (ReLU) layer. The ReLU layer may apply an element-wise 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.

[0061] 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., 16x16x12 from a 32x32x12 input volume).

[0062] One or more layers may include one or more fully connected layers. Each neuron in a fully connected layer may be connected to each neuron in a previous body. A fully connected layer may calculate a class score, and the resulting body may be 1x1x number of classes. In some examples, a CNN may include a fully connected layer, such that the output of one or more layers of the CNN may be provided as input to a fully connected layer of the CNN. In some examples, one or more convolutional streams may be implemented by one or more machine learning models 104, and some or all of the convolutional streams may include one or more fully connected layers, respectively.

[0063] In some non-limiting embodiments, one or more machine learning models 104 may include a series of convolutional layers and max pooling layers to facilitate image feature extraction, followed by multi-scale dilated convolutional layers and upsampling layers to facilitate global context feature extraction.

[0064] Although input layers, convolutional layers, pooling layers, ReLU layers, and fully connected layers are discussed herein with respect to one or more machine learning models 104, 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 in the machine learning model 104.

[0065] In embodiments where the machine learning model 104 includes a CNN, different orders and numbers of layers of the CNN may be used, depending on the embodiment. In other words, the order and number of layers of one or more machine learning models 104 are not limited to any one architecture.

[0066] In addition, some layers may include parameters (e.g., weights and / or biases), such as convolutional layers and fully connected layers, while other layers may not include parameters, such as ReLU layers and pooling layers. In some examples, the parameters may be learned by the machine learning model 104 during training. Further, some layers may include additional hyperparameters (e.g., learning rate, stride, epoch, etc.), such as convolutional layers, fully connected layers, and pooling layers, while other layers may not include, such as ReLU layers. Parameters and hyperparameters are not limited and may vary depending on the embodiment.

[0067] See now Figure 4 , each block of the method 400 described herein includes a computing process that can be performed using any combination of hardware, firmware, and / or software. For example, different functions can be performed by a processor executing instructions stored in a memory. The method 400 can also be embodied as computer usable instructions stored on a computer storage medium. The method 400 can be provided by a stand-alone application, a service or a hosted service (standalone or in combination with another hosted service), or a plug-in to another product, to name a few. In addition, by way of example, with respect to Figure 1The method 400 is described with reference to the process 100 of FIG. However, this method may additionally or alternatively be performed by any system or any combination of systems, including but not limited to the systems described herein.

[0068] Figure 4 4 is a flow chart illustrating a method 400 for training a neural network to detect intersection structure and posture according to some embodiments of the present disclosure. At block B402, the method 400 includes receiving image data representing an image depicting an intersection. For example, sensor data 102 may be received, wherein the sensor data 102 includes image data representing an image depicting an intersection.

[0069] At block B404, method 400 includes receiving first data representing an annotation corresponding to a line segment within an intersection and second data corresponding to a classification of the line segment. For example, data representing annotation 118 may be generated and / or received, wherein the annotation includes lane labels 118A corresponding to the line segment within the intersection and classification 118B of the line segment.

[0070] At block B406, method 400 includes determining the locations of key points corresponding to the line segments based at least in part on the line segments. For example, lane label 118A may include key points corresponding to a center point, a left endpoint, and / or a right endpoint of each line segment.

[0071] At block B408, the method 400 includes encoding the key points into a Gaussian heat map. For example, the encoder 120 may encode the key points into one or more heat maps 124 to be used as ground truth data 122 for training one or more machine learning models 104.

[0072] At block B410, method 400 includes training a neural network using the Gaussian heat map and classification as ground truth data. For example, one or more heat maps 124 may be used together with one or more classifications 118B as ground truth data 122 for training one or more machine learning models 104 to detect intersection structure and pose.

[0073] See now Figure 5 , Figure 5 is a data flow diagram illustrating an example process 500 for generating a path for navigating an intersection in real time or near real time using output from sensors of a vehicle according to some embodiments of the present disclosure. Figure 5 The types of intersection features described are lanes and crosswalks, but this is not intended to be limiting and is for example purposes only.

[0074] The sensor data 102 may be similar to at least Figure 1During deployment, the sensor data 102 may be applied to one or more machine learning models 104 that have been trained according to the process 100, for example, to compute heat maps 108 (and corresponding classifications), vector fields 112, and / or intensity maps 114 using the training sensor data 102. As such, the one or more machine learning models 104 may output or regress on one or more of the one or more heat maps 108, the one or more vector fields 112, and the one or more intensity maps 114, such as at least with respect to Figure 1 One or more outputs 106 are described in more detail.

[0075] Output 106 may be applied to decoder 502 to generate decoded output 504. In some examples, decoder 502 may use one or more post-processing algorithms (e.g., temporal smoothing, curve fitting, filtering, etc.) to calculate decoded output 504. For example, decoder 502 may use output 106 to generate a representation of the intersection and / or pose, which may in turn be used, for example, by path generator 516 to generate one or more paths through the intersection. Decoded output 504 may include key points 506, lane number 508, lane type 510, lane shape 512, lane heading 514, and / or other information corresponding to the intersection (e.g., line segments corresponding to the intersection that may be used to generate a visualization of the intersection for debugging, etc.).

[0076] In some embodiments, one or more of the outputs may be generated at a first spatial resolution (e.g., a downsampled spatial resolution) that is different from a second spatial resolution corresponding to the sensor data 102. As such, the decoder 502 may decode data from the first spatial resolution of the output 106 and project or attribute the data to the second spatial resolution of the sensor data 102 (e.g., using an offset vector field 112C as described herein). As a result, the processing speed of one or more machine learning models 104 may be increased while retaining computing resources for other tasks of the system. Furthermore, by increasing the processing speed, the runtime of the system may be reduced, thereby enabling the process 500 to be deployed in real time or near real time within a system (e.g., vehicle 800).

[0077] The decoder may use the heat map 108 to determine the location of the center key point, the left key point, and / or the right key point corresponding to the center of the lane, the left edge of the lane, or the right edge of the lane, respectively. For example, as described herein, the key points in the output 106 may be represented by a probability distribution of a spatial distribution within the heat map 108, and the decoder 501 may determine a 2D coordinate representing the location of the key point (e.g., at the input spatial resolution and / or the output spatial resolution). In some embodiments, because the heat map 108 may include more than one key point in a given area, the decoder 502 may apply a peak finding algorithm to determine the center location of the key point 506 as the point with the highest probability within a sliding window of a predetermined size. For example, the decoder 502 may perform a local search within a sliding window of a predetermined size, and if the probability of a pixel within the window is the highest, the pixel may be selected as the key point in the key point 506. If the center point pixel is not a peak probability pixel within the window, the peak finding algorithm may cause the center point pixel to be discarded. The size of the sliding window may be predetermined based on hyperparameter optimization performed before training one or more machine learning models 104. The decoded output 504 may then include the pixel location corresponding to the probability distribution peak achieved by non-maximum suppression as a key point in key points 506. In some embodiments, the 2D location of the surviving peak probability center key point is output by the decoder 502 as the key point 506. In a non-limiting example, the key point 506 may be filtered to remove overlap using non-maximum suppression or another filtering algorithm. In some examples, the degree of overlap between two points may be determined by calculating the distance between the two points normalized by their lengths.

[0078] The 2D position (e.g., 2D image space coordinates) of the key point 506 can then be used to determine a three-dimensional (3D) world space coordinate to assist the vehicle 800 in navigating through the world space environment. Any known method for projecting a 2D position into a 3D space can be used. In some examples, a planar model assumption can be used to project the position into a 3D space. In some other examples, LIDAR data for depth and / or distance estimation can be used for projection, and / or the future movement of the vehicle 800 as the vehicle 800 traverses the road surface can be used (e.g., to generate a more accurate ground truth for training the network), such as to determine the slope or contour of the driving surface to more accurately project the 2D point into the 3D space. In some embodiments, one or more machine learning models 104 can be trained on 3D data to directly predict the key point positions in the 3D world space.

[0079] In some embodiments, the lane type 510 (e.g., intersection entry, intersection exit, pedestrian entrance, pedestrian exit, stop line, self-lane, right adjacent lane, left adjacent lane, bicycle lane, non-drivable lane, drivable lane, combinations thereof, etc.) may be calculated by the decoder 502 using the heat map 108. The lane type 510 corresponding to the key point 506 may correspond to the label classification of the line segment as detected by the one or more machine learning models 104 (hereinafter, with respect to Figure 1 and Figure 2A-2B For example, in some non-limiting embodiments, each classification or lane label may correspond to a separate heat map 108. As such, lane type 510 may be determined directly from lane type 510 associated with the corresponding heat map. In other examples, different techniques may be used (e.g., as described herein with respect to Figure 1 The lane type 510 may be encoded in the heat map 108 using the techniques described herein. In some examples, one or more machine learning models 104 may be trained to calculate confidences corresponding to the lane type 510 and / or other semantic information, for example, using any suitable method for classification.

[0080] The number of lanes 508 and / or the number of lanes of a specific lane type 510 may be calculated by the decoder 502 by counting the number of detected key points 506 (e.g., the number of center key points) and / or by counting the number of key points corresponding to the specific lane type 510 (e.g., the center key points).

[0081] Decoder 502 may further generate lane shape 512 using directional vector field 112A and / or intensity map 114. For example, directional vector field 112A may correspond to a vector from a pixel at the output resolution to a position of a left edge keypoint, a center keypoint, and / or a right edge keypoint, as described herein. Figure 3B In more detail. In such an example, the direction vector field 112A may also be decoded by the decoder 502 to generate the lane shape 512 (e.g., in addition to using the positions of the key points 506 from the heat map 108 and / or other information from the output 106 and / or the decoded output 108). The left and right endpoints of each line segment may be determined based on the vector field 112 corresponding to the associated key points 506, and this information may be used to determine the direction or angle along the lateral dimension of the lane, the lane heading 514 (e.g., the direction perpendicular to the lateral dimension of the lane), and / or the width of the lane (e.g., the width may be the distance along the line segment, or the distance between the left edge key point and the right edge key point of a given line segment). As such, the width and / or lateral direction (e.g., the normal along the direction of travel of the lane) may be used to determine the shape of the lane 512 at least relative to the portion of the lane corresponding to the line segment and / or the left and right edge key points.

[0082] In other examples, lane shape 512 may be determined by decoding intensity map 114 to determine the width of the lane and / or decoding direction vector field 112A to determine the directionality of the line segment corresponding to the center keypoint (e.g., as described with respect to FIG. Figure 3C As described above, the lane shape 512 may be determined using the location, directionality, and width of the center keypoint (e.g., by extending a line segment from the center keypoint according to the directionality and until extending to the width of the line segment, such as extending half the width from the center keypoint along the directionality in one direction and then extending half the width from the center keypoint along the directionality in the opposite direction).

[0083] Decoder 502 may further determine lane heading 514. For example, lane heading 514 may be determined as a normal to a line segment generated between known locations of a left edge keypoint and a right edge keypoint. In other examples, lane heading 514 may be determined by decoding heading vector field 112B to determine an angle corresponding to the lane heading direction. In some examples, lane heading 514 may utilize lane type 510 and / or other semantic information output by the network, as described herein.

[0084] The decoded output 504 may be used by a path generator 516 to generate a path for navigating the intersection. The path generator 516 may connect (center) key points 506 according to their 3D real-world coordinates to generate polylines that represent potential paths for traversing the intersection by the vehicle 800 in real time or near real time. The final path may be assigned a path type determined relative to the vehicle's position, the position of the key points 506, and / or the lane heading 514 (e.g., angle). Potential path types may include, but are not limited to, turning left, turning right, switching lanes, and / or continuing in a lane.

[0085] In some examples, the path generator 516 may implement curve fitting to determine the final shape of the natural curve that most accurately reflects the potential path. Any known curve fitting algorithm can be used, such as but not limited to broken line fitting, polynomial fitting and / or convolute fitting. The shape of the potential path can be determined based on the position of the key point 506 and the corresponding heading vector (e.g., angle) associated with the key point 506 to be connected (e.g., lane heading 514). In some examples, the shape of the potential path can be aligned with the tangent of the heading vector at the position of the key point 506 to be connected. The curve fitting process can be repeated for all key points 506 that may potentially be connected to each other to generate all possible paths that the vehicle 800 may take to navigate the intersection. In some examples, based on the traffic rules and physical restrictions associated with such paths, infeasible paths can be removed from consideration. The remaining potential paths can be determined as feasible 3D paths or trajectories, and the vehicle 800 can take these 3D paths or trajectories to cross the intersection.

[0086] In some embodiments, the path generator 516 can use a matching algorithm to connect key points 506 and generate potential paths for the vehicle 800 to navigate the intersection. In these examples, a matching score can be determined for each pair of key points 506 based on the position of the key points 506, the lane heading 514 corresponding to the key points 506 (for example, two key points 506 corresponding to different directions of travel will not be connected), and the shape of the fitting curve between a pair of key points 506. Each key point 506 corresponding to the entrance of the intersection can be connected to a plurality of key points 506 corresponding to the exit of the intersection, thereby generating multiple potential paths for the vehicle 800. In some examples, a linear matching algorithm such as the Hungarian matching algorithm can be used. In other examples, a nonlinear matching algorithm such as a spectrum matching algorithm can be used to connect a pair of key points 506.

[0087] The path through the intersection can be used to perform one or more operations by the control component 518 of the vehicle 800. In some examples, the lane map can be enhanced with information related to the path. The lane map can be input to one or more control components 518 to perform different planning and control tasks. For example, the world model manager can update the world model to help navigate the intersection, the path planning layer of the autonomous driving software stack can use the intersection information to determine a path through the intersection (e.g., along one of the potential paths determined), and / or the control component can determine the control of the vehicle for navigating the intersection based on the determined path.

[0088] See now Fig. 6A , Fig. 6A The use of a neural network (e.g., Figure 5Example intersection structure prediction 600A (e.g., generated by one or more machine learning models 104) Figure 5 The prediction 600A includes the prediction corresponding to the sensor data (e.g., Figure 5 102 ). For example, lane classification 604 may correspond to an entrance of a sidewalk type, lane classification 606 may correspond to an entrance of an intersection and / or an exit of a sidewalk type, lane classification 608 may correspond to an exit of a sidewalk type, lane classification 610 may correspond to an exit of an intersection and / or an entrance of a sidewalk type, and lane classification 612 may correspond to a non-drivable lane type. Each line segment may also be associated with a center keypoint 506 and / or corresponding one or more heading directions (e.g., keypoints and associated vectors 602A-602V). As such, intersection structure and posture may be represented by a set of line segments with corresponding line classifications, keypoints, and / or heading directions.

[0089] Reference Figure 6B , Figure 6B shows the path generator (e.g., Figure 5 The path generator 516 is based on a neural network (e.g., Figure 5 Example path 600B generated by the intersection structure 600A predicted by the machine learning model 104). Paths 614A-614G are connected by Fig. 6A Paths 614A-614G may correspond to paths of the ego vehicle 800 and / or other vehicles or objects. The paths of other vehicles or objects may provide information to the vehicle 800 to help determine potential future locations of other vehicles or objects when crossing an intersection.

[0090] See now Figure 7 , each block of the method 700 described herein includes a computing process that can be performed using any combination of hardware, firmware, and / or software. For example, different functions can be performed by a processor executing instructions stored in a memory. The method 700 can also be embodied as computer usable instructions stored on a computer storage medium. To name a few, the method 700 can be provided by a stand-alone application, a service or a hosted service (standalone or in combination with another hosted service), or a plug-in for another product. In addition, by way of example, Figure 5 The method 700 is described with reference to the process 500 of FIG. However, this method may additionally or alternatively be performed by any system or any combination of systems, including but not limited to the systems described herein.

[0091] Figure 7700 for detecting and navigating intersections according to some embodiments of the present disclosure. At block B702, the method 700 includes applying image data representing an intersection in a field of view of an image sensor of a vehicle to a neural network. For example, sensor data 102 representing an image of an intersection in a field of view of an image sensor of a vehicle 800 may be applied to one or more machine learning models 104.

[0092] At block B704, the method 700 includes using a neural network and at least in part based on the image data to calculate first data representing one or more two-dimensional (2D) heat maps of locations corresponding to key points of the intersection and second data representing classification confidence values ​​corresponding to the key points. For example, the one or more machine learning models 104 can calculate one or more outputs 106 using one or more heat maps 108 including key points corresponding to the intersection and confidence values ​​corresponding to the line segment classification.

[0093] At block B706, method 700 includes decoding one or more 2D heat maps to determine pixel locations of key points in the image. For example, decoder 502 may decode heat map 108 to determine pixel locations of key points 506 in the image.

[0094] At block B708 ​​, method 700 includes determining an associated classification from the classification confidence value corresponding to the keypoint. For example, decoder 502 may determine an associated lane type 510 corresponding to keypoint 506 from the line segment classification based on the confidence value.

[0095] At block B710 , method 700 includes generating one or more suggested paths through the intersection based at least in part on the pixel locations and associated classifications. For example, path generator 516 may generate potential or suggested paths through the intersection based on the pixel locations of keypoints 506 and lane types 510 .

[0096] At block B712 , the method 700 includes performing one or more operations by the vehicle using the one or more suggested paths. For example, the one or more control components 518 may cause the vehicle 800 to perform one or more operations based on the suggested paths received from the path generator 516 .

[0097] Example Autonomous Vehicle

[0098] Fig. 8A800 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 response 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, and / or another type of vehicle (e.g., a drone and / or a vehicle accommodating one or more passengers). Autonomous vehicles are generally described according to the automation levels 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 the standard). The vehicle 800 may be capable of implementing functions consistent with one or more of Levels 3-5 of autonomous driving levels. 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).

[0099] The vehicle 800 may include components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of the vehicle. The vehicle 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. The propulsion system 850 may be connected to a drive train of the vehicle 800, which may include a transmission, to achieve propulsion of the vehicle 800. The propulsion system 850 may be controlled in response to receiving a signal from a throttle / accelerator 852.

[0100] 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.

[0101] Brake sensor system 846 may be used to operate vehicle brakes in response to receiving signals from brake actuator 848 and / or brake sensors.

[0102] May include one or more system on chip (SoC) 804 ( Figure 8C) and / or one or more controllers 836 of 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 vehicle brakes via one or more brake actuators 848, operate the steering system 854 via one or more steering actuators 856, and operate the propulsion system 850 via one or more throttles / accelerators 852. The one or more controllers 836 may include one or more onboard (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals and output operating commands (e.g., signals representing commands) to achieve autonomous driving and / or assist a human driver in driving the vehicle 800. The one or more controllers 836 may 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.

[0103] 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 input) 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.

[0104] One or more of the controllers 836 may receive inputs (e.g., represented by input data) from the instrument cluster 832 of the vehicle 800 and provide outputs (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, locations of other vehicles (e.g., 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.).

[0105] The vehicle 800 further includes a network interface 824 that can communicate via one or more networks using one or more wireless antennas 826 and / or a modem. For example, the network interface 824 may be capable of communicating via LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The one or more wireless antennas 826 may 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.

[0106] Figure 8B For use according to some embodiments of the present disclosure Fig. 8A 800. The cameras and respective fields of view are an example embodiment and are not intended to be limiting. For example, additional and / or replaceable cameras may be included and / or the cameras may be located at different locations on the vehicle 800.

[0107] The camera type 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 an 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), 820fps, 240fps, etc., depending on the embodiment. The camera may be capable of using a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In some 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.

[0108] 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).

[0109] One or more of the cameras may 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 may interfere with the camera's image data capture capabilities. With respect to the wing mirror mounting assembly, the wing mirror assembly may 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 may be integrated into the wing mirror. For side view cameras, one or more cameras may also be integrated into the four pillars at each corner of the cab.

[0110] 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 view 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.

[0111] A variety of cameras may 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 may be a wide-angle camera 870, which may be used to sense objects (e.g., pedestrians, intersection traffic, or bicycles) entering the field of view from the periphery. Figure 8B Only one wide-angle camera is shown in the figure, but there may 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 a neural network has not been trained. Long-range cameras 898 can also be used for object detection and classification and basic object tracking.

[0112] 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 each side) and an image processing chip that may measure the distance from the vehicle to the target object and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. In addition to those described herein or alternatively, other types of stereo cameras 868 may be used.

[0113] A camera having a field of view that includes a portion of the environment to the side of the vehicle 800 (e.g., a side view camera) can be used for surround viewing, providing information used to create and update the occupancy grid and generate side impact collision warnings. Figure 8B The four surround cameras 874 shown in FIG. 8 may be placed on the vehicle 800. The surround cameras 874 may include a wide-angle camera 870, a fisheye camera, a 360-degree camera, and / or the like. For example, the four fisheye cameras may be placed in front, behind, and on the sides of the vehicle. In an alternative arrangement, the vehicle may use three surround cameras 874 (e.g., left, right, and rear), and may utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround camera.

[0114] A camera having a field of view that includes a portion of the environment behind the vehicle 800 (e.g., a rear view 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 cameras as described herein (e.g., long-range and / or mid-range cameras 898, stereo cameras 868, infrared cameras 872, etc.).

[0115] Figure 8C For use according to some embodiments of the present disclosure Fig. 8ABlock diagram of an example system architecture of an example autonomous vehicle 800. It should 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 combination with other components, and in any appropriate combination and location. The various functions described herein as being performed by an entity may be implemented by hardware, firmware, and / or software. For example, the various functions may be implemented by a processor executing instructions stored in a memory.

[0116] Figure 8C Each of the components, features, and systems of the vehicle 800 is illustrated as being connected via a bus 802. The bus 802 may include a controller area network (CAN) data interface (alternatively, referred to herein as a "CAN bus"). The CAN may be a network inside the vehicle 800 that assists in controlling various features and functions of the vehicle 800, such as the 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 compatible.

[0117] Although bus 802 is described as a CAN bus here, this is not intended to be limiting. For example, in addition to or alternatively to the CAN bus, FlexRay and / or Ethernet can be used. In addition, although bus 802 is represented by a single line, this is not intended to be limiting. For example, there can 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 can be used to perform different functions, and / or can be used for redundancy. For example, a first bus 802 can be used for a collision avoidance function, and a second bus 802 can be used for drive control. In any example, each bus 802 can communicate with any component of the vehicle 800, and two or more buses 802 can communicate with the same component. In some examples, each SoC 804, each controller 836, and / or each computer in the vehicle can have access to the same input data (e.g., input from a sensor of the vehicle 800), and can be connected to a common bus such as a CAN bus.

[0118] The vehicle 800 may include one or more controllers 836, such as those described herein. Fig. 8A Those controllers described. Controller 836 can be used for a variety of functions. Controller 836 can be coupled to any other different components and systems of vehicle 800, and can be used for control of vehicle 800, artificial intelligence of vehicle 800, infotainment for vehicle 800, and / or the like.

[0119] The vehicle 800 may include one or more system on chip (SoC) 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, one or more SoCs 804 may be combined with an HD map 822 in a system (e.g., a system of the vehicle 800), and the HD map may be downloaded from one or more servers (e.g., a server) via a network interface 824. Fig.8D one or more servers 878) to obtain map refreshes and / or updates.

[0120] 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 of which has 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 clusters of the CPU 806 can be active at any given time.

[0121] 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 save 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. CPU 806 may further implement an enhanced algorithm for managing power states, in which allowed power states and expected wake-up 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 the work being offloaded to the microcode.

[0122] 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 storage capacity), and two or more of these streaming microprocessors may share an L2 cache (e.g., an L2 cache with 512KB 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).

[0123] 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, thread beam (warp) scheduler, dispatch unit and / or 64KB register file. In addition, the streaming microprocessor can include independent parallel integer and floating point data paths to provide efficient execution of workloads using a mix of calculations and addressing calculations. The streaming microprocessor may include independent thread scheduling capabilities to allow for finer-grained synchronization and collaboration between parallel threads. A streaming microprocessor may include a combined L1 data cache and shared memory unit to increase performance while simplifying programming.

[0124] The GPU 808 may include a 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), may be used in addition to or in lieu of HBM memory.

[0125] 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 service (ATS) support can be used to allow the GPU 808 to directly access the CPU 806 page table. In such an example, when the GPU 808 memory management unit (MMU) experiences a miss, an address translation request can be transmitted to the CPU 806. In response, the CPU 806 can look up the virtual-physical mapping for the address in its page table and transmit the translation back to the GPU 808. In this way, the unified memory technology can allow a single unified virtual address space to be used for the memory of both the CPU 806 and the GPU 808, thereby simplifying GPU 808 programming and porting applications to the GPU 808.

[0126] In addition, GPU 808 may include access counters that can track how often GPU 808 accesses the memory of other processors. The access counters can help ensure that memory pages are moved to the physical memory of the processor that accesses them most frequently.

[0127] 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 may track the state of a line, for example, by using a cache coherence protocol (e.g., MEI, MESI, MSI, etc.). Depending on the embodiment, the L3 cache may include 4MB or more, although smaller cache sizes may also be used.

[0128] SoC 804 may include one or more accelerators 814 (e.g., hardware accelerators, software accelerators, or combinations thereof). For example, SoC 804 may include a hardware acceleration cluster, which may include optimized hardware accelerators and / or large on-chip memory. The large on-chip memory (e.g., 4MB SRAM) may enable the hardware acceleration cluster to accelerate neural networks and other calculations. The hardware acceleration cluster may be used to supplement GPU 808 and offload some tasks of GPU 808 (e.g., free up more cycles of GPU 808 for performing other tasks). As an example, accelerator 814 may be used for targeted workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are sufficiently stable to easily control 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).

[0129] The accelerator 814 (e.g., a hardware acceleration cluster) may include a deep learning accelerator (DLA). The DLA may include one or more tensor processing units (TPUs) that may 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 exceeds the performance of the CPU. The TPU may perform several functions, including a single instance convolution function, support, for example, INT8, INT16, and FP16 data types for both features and weights, and post-processor functions.

[0130] The 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: a CNN for object recognition and detection using data from a camera sensor; a CNN for distance estimation using data from a camera sensor; a CNN for emergency vehicle detection and identification and detection using data from a microphone; a CNN for facial recognition and vehicle owner identification using data from a camera sensor; and / or a CNN for safety and / or security-related events.

[0131] 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.

[0132] The accelerator 814 (e.g., a hardware acceleration 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.

[0133] 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 number of memories. 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 storage devices. For example, the RISC core can include an instruction cache and / or a tightly coupled RAM.

[0134] The DMA may enable components of the PVA to access system memory independently of the CPU 806. The DMA may 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 may support addressing in up to six or more dimensions, which may include block width, block height, block depth, horizontal block stepping, vertical block stepping, and / or depth stepping.

[0135] The vector processor can be a programmable processor that can be designed to efficiently and flexibly perform 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 rate.

[0136] Each of the vector processors may include an instruction cache and may be coupled to a dedicated memory. As a result, in some examples, each of the vector processors may be configured to execute independently of the other vector processors. In other examples, the vector processors included in a particular PVA may be configured to employ data parallelism. For example, in some embodiments, multiple vector processors included in a single PVA may execute the same computer vision algorithm, but on different regions of an image. In other examples, the vector processors included in a particular PVA may execute different computer vision algorithms simultaneously on the same image, or even execute different algorithms on sequence images or portions of an image. Among other things, any number of PVAs may be included in a hardware acceleration cluster, and any number of vector processors may be included in each of these PVAs. In addition, the PVA may include additional error correction code (ECC) memory to enhance overall system security.

[0137] The accelerator 814 (e.g., a hardware acceleration 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 not limited to, 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).

[0138] The on-chip computer vision network may include an interface that determines that both the PVA and DLA provide ready and valid signals before transmitting any control signals / addresses / data. Such an interface may provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst communication for continuous data transmission. This type of interface may comply with ISO 26262 or IEC 61508 standards, but other standards and protocols may also be used.

[0139] In some examples, SoC 804 may include a real-time ray tracing hardware accelerator such as described in U.S. Patent Application No. 16 / 101,232 filed on August 10, 2018. The real-time ray tracing hardware accelerator may be used to quickly and efficiently determine the position and range of objects (e.g., within a world model) in order to generate real-time visualization simulations for RADAR signal interpretation, for sound propagation synthesis and / or analysis, for SONAR system simulations, for general wave propagation simulations, for comparison with LIDAR data for the purpose of positioning and / or other functions, and / or for other uses.

[0140] The accelerator 814 (e.g., a cluster of hardware accelerators) has a wide range of uses in autonomous driving. The PVA can be a programmable visual 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-dense or dense rule calculations, 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 effective at object detection and integer math.

[0141] For example, according to one embodiment of the technology, PVA is used to perform computer stereo vision. In some examples, algorithms based on semi-global matching can be used, but this is not intended to be limiting. Many applications for level 3-5 autonomous driving require instant 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.

[0142] In some examples, PVA can be used to perform dense optical flow. According to the process raw RADAR data (e.g., using 4D fast Fourier transform) to provide processed RADAR. In other examples, PVA is used for time-of-flight depth processing, which is for example by processing raw time-of-flight data to provide processed time-of-flight data.

[0143] DLA can be used to run any type of network to enhance control and driving safety, including, for example, a neural network that outputs a confidence measure for each object detection. Such a confidence value can be interpreted as a probability, or 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 positive detections rather than false positive detections. For example, the system can set a threshold for confidence and only consider detections that exceed the threshold as true positive detections. In an automatic emergency braking (AEB) system, a false positive detection will cause the vehicle to automatically perform emergency braking, which is obviously undesirable. Therefore, only the most confident detection should be considered a trigger for AEB. DLA can run a neural network for regressing confidence values. The neural network can take at least some subset of parameters as its input, such as bounding box dimensions, ground plane estimates obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 866 output related to vehicle 800 orientation and distance, 3D position estimates of objects obtained from neural networks and / or other sensors (e.g., LIDAR sensor 864 or RADAR sensor 860), etc.

[0144] SoC 804 may include one or more data stores 816 (e.g., memory). Data store 816 may be on-chip memory of SoC 804 that may store neural networks to be executed on the GPU and / or DLA. In some examples, data store 816 may be large enough to store multiple instances of the neural network for redundancy and safety. Data store 812 may include an L2 or L3 cache 812. References to data store 816 may include references to memory associated with PVA, DLA, and / or other accelerators 814 as described herein.

[0145] SoC 804 may include one or more processors 810 (e.g., embedded processors). Processor 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 the temperature, and SoC 804 may use the ring oscillator to detect the temperature of 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 SoC 804 in a lower power state and / or place the vehicle 800 in a driver safety parking mode (e.g., parking the vehicle 800 safely).

[0146] The processor 810 may further include a set of embedded processors that may be used as an audio processing engine. The audio processing engine may be an audio subsystem that 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.

[0147] The processor 810 may further include an always-on processor engine that may provide the necessary hardware features to support low-power sensor management and wake-up use cases. The always-on processor engine may include a processor core, tightly coupled RAM, supporting peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0148] The processor 810 may further 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 (e.g., timers, interrupt controllers, etc.), and / or routing logic. In safety mode, the two or more cores may operate in lockstep mode and act as a single core with comparison logic to detect any differences between their operations.

[0149] Processor 810 may further include a real-time camera engine, which may include a dedicated processor subsystem for handling real-time camera management.

[0150] Processor 810 may further include a high dynamic range signal processor, which may include an image signal processor, which is a hardware engine that is part of the camera processing pipeline.

[0151] The processor 810 may include a video image compositer, which may be a processing block (e.g., implemented on a microprocessor) that implements video post-processing functions required for the video playback application to produce a 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 an in-cab surveillance camera sensor. The in-cab surveillance 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.

[0152] The video image compositer may include enhanced temporal noise reduction for both spatial and temporal noise reduction. For example, in the case of motion in the video, the noise reduction appropriately weights the spatial information and reduces the weight of information provided by adjacent frames. In the case where an image or portion of an image does not include motion, the temporal noise reduction performed by the video image compositer may use information from previous images to reduce noise in the current image.

[0153] The video image compositor may also be configured to perform stereoscopic rectification on the input stereoscopic footage frames. The video image compositor may 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 may be used to offload the GPU 808 to improve performance and responsiveness.

[0154] The SoC 804 may further include a mobile industry processor interface (MIPI) camera serial interface for receiving video and input from a camera, a high-speed interface, and / or a video input block that may be used for camera and related pixel input functions. The SoC 804 may further include an input / output controller that may be controlled by software and may be used to receive I / O signals that are not committed to a specific role.

[0155] SoC 804 may further include a wide range of peripheral interfaces to enable communication with peripherals, audio codecs, power management and / or other devices. 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 bus 802 (e.g., speed of vehicle 800, steering wheel position, etc.), data from GNSS sensor 858 (connected via Ethernet or CAN bus). SoC 804 may further include dedicated high-performance mass storage controllers, which may include their own DMA engines, and which may be used to free up CPU 806 from routine data management tasks.

[0156] SoC 804 can be an end-to-end platform with a flexible architecture that spans automation levels 3-5, providing a comprehensive functional safety architecture that leverages and efficiently uses computer vision and ADAS technologies to achieve diversity and redundancy, together 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 efficient 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.

[0157] The technology thus provides capabilities and functionality that cannot be achieved by 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, for example, execution time and power consumption. In particular, many CPUs are unable to execute complex object detection algorithms in real time, which is a requirement for in-vehicle ADAS applications and a requirement for practical Level 3-5 autonomous vehicles.

[0158] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a hardware acceleration cluster, the techniques described herein allow multiple neural networks to be executed simultaneously and / or sequentially, and the results to be combined together to achieve Level 3-5 autonomous driving functions. For example, a CNN executed on a DLA or dGPU (e.g., GPU 820) may 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 may further include a neural network capable of recognizing, interpreting, and providing semantic understanding of the signs, and passing that semantic understanding to a path planning module running on the CPU complex.

[0159] As another example, as required for Level 3, 4, or 5 driving, multiple neural networks may be running simultaneously. For example, a warning sign consisting of "Caution: Flashing Lights Indicate Icing Conditions" along with electric lights may be interpreted by several neural networks independently or collectively. The sign itself may be recognized as a traffic sign by a deployed first neural network (e.g., a trained neural network), and the text "Flashing Lights Indicate Icing Conditions" may be interpreted by a deployed second neural network that 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 may be identified by operating a deployed third neural network over multiple frames that informs the vehicle's path planning software of the presence (or absence) of the flashing lights. All three neural networks may be running simultaneously, for example, within the DLA and / or on the GPU 808.

[0160] In some examples, a CNN for facial recognition and owner recognition can use data from the 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.

[0161] 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 rate of emergency vehicles (e.g., by using the Doppler effect). The CNN can also be trained to identify emergency vehicles specific to the local area in which the vehicle is operating, as identified by 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 emergency vehicle safety routines to slow the vehicle, drive to the side of the road, stop the vehicle, and / or idle the vehicle until the emergency vehicle passes.

[0162] 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.

[0163] 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 on inputs (e.g., sensor data) from sensors of the vehicle 800.

[0164] The vehicle 800 may further 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 cellular antennas, Bluetooth antennas, etc.). The network interface 824 can be used to enable wireless connections to the cloud (e.g., with 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 collaborative adaptive cruise control functionality of the vehicle 800.

[0165] 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 a radio frequency front end for up-conversion from baseband to radio frequency and down-conversion from radio frequency to baseband. The frequency conversion may be performed by a known process and / or may be performed using a super-heterodyne process. In some examples, the radio frequency front end function may be provided by a separate chip. The network interface may include a wireless function for communicating via LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0166] The vehicle 800 may further include a data store 828, which may include off-chip storage (e.g., outside the SoC 804). The data store 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.

[0167] The vehicle 800 may further include a GNSS sensor 858. The GNSS sensor 858 (e.g., GPS and / or assisted GPS sensor) is used to assist in 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.

[0168] The vehicle 800 may further include a RADAR sensor 860. The RADAR sensor 860 may be used by the vehicle 800 for remote vehicle detection even in darkness and / or inclement weather conditions. The RADAR functional safety level may be ASIL B. The RADAR sensor 860 may 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, accessing Ethernet in some examples to access raw data. A variety of RADAR sensor types may be used. For example and without limitation, the RADAR sensor 860 may be suitable for front, rear, and side RADAR use. In some examples, a pulse Doppler RADAR sensor is used.

[0169] The RADAR sensor 860 may 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 the like. In some examples, the long-range RADAR may be used for adaptive cruise control functions. The long-range RADAR system may provide a wide field of view (e.g., within a range of 250m) achieved by two or more independent scans. The RADAR sensor 860 may help distinguish between static objects and moving objects, and may 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 an example with six antennas, the central four antennas may create a focused beam pattern designed to record the surroundings of the vehicle 800 at a higher rate with minimal traffic interference from adjacent lanes. The other two antennas may expand the field of view, making it possible to quickly detect vehicles entering or leaving the lane of the vehicle 800.

[0170] 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 a rear bumper. When mounted on both ends of a rear bumper, such a RADAR sensor system may create two beams that continuously monitor the blind spots behind and beside the vehicle.

[0171] Short-range RADAR systems can be used in ADAS systems for blind spot detection and / or lane change assistance.

[0172] The vehicle 800 may further include ultrasonic sensors 862. The ultrasonic sensors 862, which may be placed on the front, rear, and / or sides of the vehicle 800, may be used for parking assistance and / or creating and updating occupancy grids. A variety of ultrasonic sensors 862 may be used, and different ultrasonic sensors 862 may be used for different detection ranges (e.g., 2.5 m, 4 m). The ultrasonic sensors 862 may operate at a functional safety level of ASIL B.

[0173] The vehicle 800 may include a LIDAR sensor 864. The LIDAR sensor 864 may be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensor 864 may be ASIL B for functional safety level. In some examples, the 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).

[0174] In some examples, the LIDAR sensor 864 may be able to provide 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 800m, an accuracy of 2cm-3cm, and support for 800Mbps Ethernet connections. In some examples, one or more non-protruding 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, side, and / or corner of the vehicle 800. In such examples, the LIDAR sensor 864 may provide a field of view of up to 820 degrees horizontally and 35 degrees vertically, with a range of 200m, 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.

[0175] 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 200m. 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 highly accurate and distortion-free images of the surrounding environment to be generated with each laser flash. In some examples, four flash LIDAR sensors may 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 fans. 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.

[0176] The vehicle may further 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, for example, in a six-axis application, the IMU sensor 866 may include an accelerometer and a gyroscope, and in a nine-axis application, the IMU sensor 866 may include an accelerometer, a gyroscope, and a magnetometer.

[0177] In some embodiments, the IMU sensor 866 can be implemented as a miniature high-performance GPS-assisted inertial navigation system (GPS / INS) that combines a micro-electromechanical system (MEMS) inertial sensor, a high-sensitivity GPS receiver, and an advanced Kalman filter algorithm to provide estimates of position, velocity, and attitude. Thus, in some examples, the IMU sensor 866 can enable the vehicle 800 to estimate heading by directly observing and correlating velocity changes from the GPS to the IMU sensor 866 without the need for input from a magnetic sensor. In some examples, the IMU sensor 866 and the GNSS sensor 858 can be combined into a single integrated unit.

[0178] 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.

[0179] The vehicle may further include any number of camera types, including a stereo camera 868, a wide angle camera 870, an infrared camera 872, a surround camera 874, a long-range and / or mid-range camera 898, and / or other camera types. These cameras can 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 can be used to provide the necessary coverage around the vehicle 800. In addition, the number of cameras can 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 can support Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Each of the cameras described herein may be capable of supporting Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Fig. 8A and Figure 8B Described in more detail.

[0180] The vehicle 800 may further include a vibration sensor 842. The vibration sensor 842 may measure vibrations of components of the vehicle, such as axles. For example, changes in vibration may indicate changes in the road surface. In another example, when two or more vibration sensors 842 are used, the difference between the vibrations may be used to determine friction or slip of the road surface (e.g., when there is a vibration difference between a powered drive shaft and a free-spinning shaft).

[0181] 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.

[0182] The ACC system may use a RADAR sensor 860, a LIDAR sensor 864, and / or a camera. The ACC system may include a longitudinal ACC and / or a lateral ACC. The longitudinal ACC monitors and controls the distance to the vehicle immediately in front of the vehicle 800, and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle in front. The lateral ACC performs distance keeping and suggests that the vehicle 800 change lanes when necessary. The lateral ACC is related to other ADAS applications such as LCA and CWS.

[0183] 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 a network interface 824 and / or a 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 farther ahead. The CACC system may 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 is possible to improve the smoothness of traffic flow and reduce road congestion.

[0184] 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 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 components. The FCW system can provide warnings in the form of, for example, sound, visual warnings, vibrations and / or rapid brake pulses.

[0185] 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 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 may include technologies such as dynamic brake support and / or collision approach braking.

[0186] 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 not activated. The LDW system may use a front-facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component.

[0187] The LKA system is a variation of the LDW system. If the vehicle 800 begins to leave the lane, the LKA system provides steering input or braking to correct the vehicle 800.

[0188] The BSW system detects and warns the driver of vehicles in the car's blind spot. The BSW system can provide visual, auditory and / or tactile alerts to indicate that it is unsafe to merge or change lanes. 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, which is electrically coupled to driver feedback such as a display, speaker and / or vibration components.

[0189] The RCTW system may provide visual, audible, and / or tactile notifications when an object is detected outside the range of the rear camera while the vehicle 800 is in reverse. Some RCTW systems include AEB to ensure that the vehicle brakes are applied to avoid a crash. The RCTW system may use one or more rear 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 vibrating component.

[0190] Conventional ADAS systems may be prone to false positive results, which may annoy and distract the driver, but are typically not catastrophic because the ADAS system alerts the driver and allows the driver to decide whether the safety condition really exists and take action accordingly. However, in the autonomous vehicle 800, in the case 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 faults 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 coordinate the conflict to ensure safe operation.

[0191] In some examples, the master computer can be configured to provide a confidence score to the supervisory MCU, indicating the master computer's confidence in the selected result. If the confidence score exceeds a threshold, the supervisory MCU can follow the direction of the master computer, regardless of whether the auxiliary computer provides conflicting or inconsistent results. In the case where the confidence score does not meet the threshold and the master computer and the auxiliary computer indicate different results (e.g., conflicts), the supervisory MCU can arbitrate between these computers to determine the appropriate result.

[0192] The supervisory MCU may be configured to run a neural network that is trained and configured to determine conditions under which the auxiliary computer provides a false alarm based on outputs from the primary computer and the auxiliary computer. Thus, the neural network in the supervisory MCU may learn when the output of the auxiliary computer may be trusted and when it may not. For example, when the auxiliary computer is a RADAR-based FCW system, the neural network in the supervisory MCU may learn when the FCW system is identifying a metal object that is not actually dangerous, such as a drainage grate or manhole cover that triggers an alarm. Similarly, when the auxiliary computer is a camera-based LDW system, the neural network in the supervisory MCU may learn to ignore the LDW when a cyclist or pedestrian is present and lane departure is actually the safest strategy. In an embodiment that includes a neural network running on the supervisory MCU, the supervisory MCU may include at least one of a DLA or GPU suitable for running the neural network with associated memory. In a preferred embodiment, the supervisory MCU may include a component of the SoC 804 and / or be included as a component of the SoC 804.

[0193] 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 for failures caused by software (or software-hardware interface) functions. For example, if there is a software vulnerability or error 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 does not cause a substantial error.

[0194] In some examples, the output of the ADAS system 838 can be fed to the perception block of the main computer and / or the dynamic driving task block of the main computer. For example, if the ADAS system 838 indicates a forward collision warning due to an object immediately ahead, the perception block can use this information in identifying the object. In other examples, the auxiliary computer can have its own neural network that is trained and thus reduces the risk of false positives as described herein.

[0195] The vehicle 800 may further 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 discrete components. The infotainment SoC 830 may include a combination of hardware and software that may be used to provide audio (e.g., music, personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., TV, movies, streaming, etc.), phone (e.g., hands-free calling), network connectivity (e.g., LTE, Wi-Fi, etc.), and / or information services (e.g., 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, onboard entertainment, Wi-Fi, 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.

[0196] 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 master controller 836 (e.g., a main and / or backup computer of the vehicle 800), the GPU of the infotainment system 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.

[0197] The vehicle 800 may further 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 discrete controller or a supercomputer). The instrument cluster 832 may include a set of instruments, such as a speedometer, a fuel level, an oil pressure, a tachometer, an odometer, a turn indicator, a shift position indicator, a seat belt warning light, a parking brake warning light, an engine fault light, an airbag (SRS) system information, lighting controls, safety system controls, navigation information, and the like. In some examples, information may 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.

[0198] Fig.8D For a cloud-based server and Fig. 8A 800. System 876 may include a server 878, a network 890, and a vehicle including the vehicle 800. The server 878 may include a plurality of GPUs 884(A)-884(H) (collectively referred to herein as GPUs 884), PCIe switches 882(A)-882(H) (collectively referred to herein as PCIe switches 882), and / or CPUs 880(A)-880(B) (collectively referred to herein as CPUs 880). The GPUs 884, CPUs 880, and PCIe switches may be interconnected with high-speed interconnects and / or PCIe connections 886 such as, for example and without limitation, the NVLink interface 888 developed by NVIDIA. In some examples, the GPUs 884 are connected via NVLink and / or NVSwitch SoCs, and the GPUs 884 and PCIe switches 882 are connected via PCIe interconnects. Although eight GPUs 884, two CPUs 880, and two PCIe switches are illustrated, this is not intended to be limiting. Depending on the embodiment, each of the servers 878 may include any number of GPUs 884, CPUs 880, and / or PCIe switches. For example, each of the servers 878 may include eight, sixteen, thirty-two, and / or more GPUs 884.

[0199] Server 878 may receive image data over network 890 and from a vehicle, the image data representing images showing unexpected or changed road conditions, such as recently begun road work. Server 878 may transmit neural network 892, updated neural network 892, and / or map information 894, including information about traffic and road conditions, over network 890 and to the vehicle. Updates to map information 894 may 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 may have been generated from new training and / or data received from any number of vehicles in the environment and / or based on experience from training performed at a data center (e.g., using server 878 and / or other servers).

[0200] 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 a simulation (e.g., using a game engine). In some examples, the training data is labeled (e.g., when 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., when 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: categories such as supervised training, semi-supervised training, unsupervised training, self-learning, reinforcement learning, joint learning, transfer learning, feature learning (including main components and cluster analysis), multilinear subspace learning, manifold learning, representation learning (including alternate 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.

[0201] 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 may include a deep learning supercomputer and / or a dedicated AI computer powered by GPU 884, such as DGX and DGX Station machines developed by NVIDIA. However, in some examples, server 878 may include a deep learning infrastructure of a data center powered only by CPUs.

[0202] 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 a sequence of images and / or objects located in the sequence of images that vehicle 800 has located (e.g., via computer vision and / or other machine learning object classification techniques). The deep learning infrastructure may run its own neural network to identify objects and compare them to the objects identified by vehicle 800, and if the results do not match and the infrastructure concludes that the AI ​​in vehicle 800 has failed, then server 878 may transmit a signal to vehicle 800 instructing the fail-safe computer of vehicle 800 to take control, notify passengers, and complete a safe parking maneuver.

[0203] For reasoning, the server 878 may include a GPU 884 and one or more programmable reasoning accelerators (e.g., NVIDIA's TensorRT). The combination of GPU-powered servers and reasoning acceleration can make real-time responses possible. In other examples, such as when performance is not so important, CPU, FPGA, and other processor-powered servers can be used for reasoning.

[0204] Example computing device

[0205] Fig. 9 900 is a block diagram of an example computing device 900 suitable for implementing some embodiments of the present disclosure. The computing device 900 may include a bus 902 that directly or indirectly couples the following devices: a memory 904, one or more central processing units (CPUs) 906, one or more graphics processing units (GPUs) 908, a communication interface 910, an input / output (I / O) port 912, an input / output component 914, a power supply 916, and one or more presentation components 918 (e.g., a display).

[0206] although Fig. 9 The various blocks of are shown as being connected via a bus 902 having wires, but this is not intended to be limiting and is merely for clarity. For example, in some embodiments, a presentation component 918 such as a display device may be considered an I / O component 914 (e.g., if the display is a touch screen). As another example, the CPU 906 and / or the GPU 908 may include a memory (e.g., the memory 904 may represent a storage device other than the memory of the GPU 908, the CPU 906, and / or other components). In other words, Fig. 9The computing devices described herein are illustrative only. No distinction is made between categories such as "workstations," "servers," "laptops," "desktops," "tablets," "client devices," "mobile devices," "handheld devices," "game consoles," "electronic control units (ECUs)," "virtual reality systems," and / or other device or system types, as all are considered Fig. 9 within the range of computing devices.

[0207] The bus 902 may represent one or more buses, such as an address bus, a data bus, a control bus, or a combination thereof. The bus 902 may include one or more bus types, such as an Industry Standard Architecture (ISA) bus, an Extended Industry Standard Architecture (EISA) bus, a Video Electronics Standards Association (VESA) bus, a Peripheral Component Interconnect (PCI) bus, a Peripheral Component Interconnect Express (PCIe) bus, and / or another type of bus.

[0208] The memory 904 may include any of a variety of computer-readable media. Computer-readable media may be any available media that can be accessed by the computing device 900. Computer-readable media may include volatile and non-volatile media and removable and non-removable media. By way of example and not limitation, computer-readable media may include computer storage media and communication media.

[0209] Computer storage media may include volatile and nonvolatile media and / or removable and non-removable media implemented in any method or technology for storing 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 programs and / or program elements, such as an operating system). Computer storage media may include, but are not limited to, RAM, ROM, EEPROM, flash memory or other storage technology, CD-ROM, digital versatile disk (DVD) or other optical disk storage, cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium that can be used to store the desired information and can be accessed by the computing device 900. When used herein, computer storage media does not include the signal itself.

[0210] Communication media may include computer readable instructions, data structures, program modules, and / or other data types in a modulated data signal such as a carrier wave or other transmission mechanism, and include any information transport medium. The term "modulated data signal" may refer to a signal that has one or more of its characteristics set or changed in such a way that information is encoded into the signal. By way of example and not limitation, communication media may include wired media such as a wired network or a direct wired connection, and wireless media such as sound, RF, infrared, and other wireless media. Any combination of the above should also be included within the scope of computer readable media.

[0211] The CPU 906 may be configured to execute 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. Each of the CPUs 906 may include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of simultaneously processing a large number of software threads. The CPU 906 may include any type of processor, and may include different types of processors, depending on the type of computing device 900 implemented (e.g., a processor with fewer cores for mobile devices and a processor with more cores for servers). For example, depending on the type of computing device 900, the processor may be an ARM processor implemented using reduced instruction set computing (RISC) or an x86 processor implemented using complex instruction set computing (CISC). In addition to one or more microprocessors or supplementary coprocessors such as mathematical coprocessors, the computing device 900 may also include one or more CPUs 906.

[0212] GPU 908 can be used by computing device 900 to render graphics (e.g., 3D graphics). GPU 908 can include hundreds or thousands of cores capable of processing hundreds or thousands of software threads simultaneously. GPU 908 can generate pixel data for outputting an image in response to a rendering command (e.g., a rendering command from CPU 906 received via a host interface). GPU 908 can include a graphics memory such as a display memory for storing pixel data. Display memory can be included as part of memory 904. GPU 908 can include two or more GPUs operating in parallel (e.g., via a link). When combined together, each GPU 908 can generate pixel data for different parts of an output image or for different output images (e.g., a first GPU for a first image, a second GPU for a second image). Each GPU can include its own memory, or can share memory with other GPUs.

[0213] In examples where computing device 900 does not include GPU 908, CPU 906 may be used to render graphics.

[0214] The communication interface 910 may include one or more receivers, transmitters, and / or transceivers that enable the computing device 700 to communicate with other computing devices via an electronic communication network, including wired and / or wireless communications. The communication interface 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., communicating over Ethernet), a low power wide area network (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0215] The I / O ports 912 can enable the computing device 900 to be logically coupled to other devices including I / O components 914, 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 microphones, mice, keyboards, joysticks, game pads, game controllers, satellite dishes, scanners, printers, wireless devices, and the like. The I / O components 914 can provide a natural user interface (NUI) that processes air gestures, voice, or other physiological inputs generated by a user. In some instances, 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 adjacent to the screen, air gestures, head and eye tracking, and touch recognition associated with the display of the computing device 900 (as described in more detail below). The computing device 900 may include a depth camera such as a stereo camera system, an infrared camera system, an RGB camera system, touch screen technology, and combinations of these for gesture detection and recognition. In addition, computing device 900 may include an accelerometer or gyroscope to enable motion detection (e.g., as part of an inertial measurement unit (IMU)). In some examples, the output of the accelerometer or gyroscope may be used by computing device 900 to render immersive augmented reality or virtual reality.

[0216] 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.

[0217] 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), a speaker, 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 an image, video, sound, etc.).

[0218] The present disclosure may be described in the general context of machine-usable instructions or computer codes 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. Typically, program modules including routines, programs, objects, components, data structures, etc. refer to codes that perform specific tasks or implement specific abstract data types. The present disclosure may be practiced in a variety of system configurations, including handheld devices, consumer electronics, general-purpose computers, more specialized computing devices, etc. The present disclosure may also be practiced in a distributed computing environment in which tasks are performed by remote processing devices linked through a communication network.

[0219] As used herein, the description of "and / or" about two or more elements should be interpreted as referring to only one element or a combination of elements. For example, "element A, element B and / or element C" may include only element A, only element B, only element C, element A and element B, element A and element C, element B and element C, or elements A, B and C. In addition, "at least one of element A or element B" may include at least one of element A, at least one of element B, or at least one of element A and at least one of element B. 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.

[0220] 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 inventors have contemplated that the claimed subject matter may also be embodied in other ways to include steps 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 method for detecting intersection posture, include: applying image data representing an intersection in a field of view of an image sensor of the vehicle to one or more neural networks; Calculating first data and second data using the one or more neural networks and based at least on the image data, the first data representing one or more two-dimensional (2D) heat maps representing locations of key points on a line segment corresponding to the intersection, and the second data representing classification confidence values ​​corresponding to the key points; decoding the one or more 2D heat maps to determine pixel locations of the keypoints in the image; determining an associated classification corresponding to the key point from the classification confidence value; connecting at least two of the key points based at least on the pixel locations and the associated classifications to generate a suggested path through the intersection; as well as One or more operations are performed by the vehicle using the suggested path.

2. The method according to claim 1, wherein the associated classification represents one or more of the following items: an entry point of a crosswalk, an entry point of the intersection, an exit point of the intersection, an exit point of a crosswalk, or a bicycle lane.

3. The method of claim 1, wherein the 2D heat map comprises an unnormalized normal distribution, and wherein the 2D heat map comprises one of a fixed standard deviation or an adaptive covariance.

4. The method according to claim 1, in, Decoding the 2D heat map includes executing a peak finding algorithm to determine peak probability pixels, the peak probability pixels corresponding to the key points.

5. The method according to claim 1, further comprising: include: Third data corresponding to a suggested maneuver of the vehicle through the intersection is received, wherein the one or more operations include selecting the suggested path based at least on the third data.

6. The method according to claim 1, further comprising: include: calculating, using the one or more neural networks, third data representing one or more vector fields corresponding to one or more of a directionality of a line segment corresponding to the key point or a heading direction of a lane corresponding to the key point; and decoding the one or more vector fields based at least on the third data and the pixel position of the key point to determine one or more of the directionality or the heading direction. 7 . The method of claim 6 , wherein generating the suggested path is further based at least on one or more of the directionality or the heading direction.

8. The method according to claim 1, further comprising: include: calculating, using the one or more neural networks, third data representing an intensity map corresponding to a width of a lane associated with the intersection; and decoding the intensity map to determine the width of the lane based at least on the third data and the pixel positions of the key points.

9. The method of claim 1, wherein the suggested path includes at least one of a left turn, a right turn, or a straight path for the vehicle to traverse the intersection.

10. The method according to claim 1, further comprising: include: Third data representing an offset vector field is calculated using the one or more neural networks, wherein decoding the 2D heat map is also based on at least the offset vector field.

11. The method of claim 1, wherein generating the suggested path include: A curve fitting algorithm is performed to determine a trajectory from a first one of the keypoints to a second one of the keypoints.

12. A method for detecting intersection posture, include: Receiving image data representing an intersection; receiving first data representing annotations corresponding to line segments within the intersection and second data corresponding to classifications of the line segments; determining locations of key points corresponding to the line segments based at least on the line segments, wherein one or more suggested paths through the intersection are generated based at least on the first data, the second data, and at least one pair of connectable key points; encoding the key points into a Gaussian heat map; and training one or more neural networks using the Gaussian heat map and the classifications as ground truth data.

13. The method according to claim 12, wherein determining the position of the key point include: The pixel corresponding to the longitudinal center of the line segment is identified.

14. The method according to claim 13, further comprising: include: encoding one or more vector fields corresponding to one or more of a directionality of the line segment or a heading direction of a lane corresponding to the line segment, wherein training the one or more neural networks further uses the one or more vector fields.

15. The method according to claim 12, further comprising: include: The vector field is encoded at least by generating a vector from a pixel at an output resolution of the one or more neural networks to a first keypoint location within the output resolution, the first keypoint location corresponding to a second keypoint location at an input resolution of the one or more neural networks, wherein the first keypoint location and the second keypoint location correspond to at least one of a center keypoint, a left edge keypoint, or a right edge keypoint of the line segment, and further, wherein training the one or more neural networks further uses the vector field.

16. The method according to claim 12, further comprising: include: The intensity map is encoded based on at least a length of the line segment, the length corresponding to a width of a lane within the intersection, and further wherein training the one or more neural networks further uses the intensity map.

17. A detection system for intersection posture, include: an image sensor for generating image data representing an image of a field of view of the image sensor in an environment; A computing device comprising one or more processing devices and one or more memory devices communicatively coupled to the one or more processing devices, the one or more memory devices having stored thereon programming instructions that, when executed by the processor, cause the following to be instantiated: applying image data representing an image of an intersection in a field of view of an image sensor of a vehicle to one or more neural networks; an output generator for computing, using one or more neural networks and based at least on the image data, first data representing one or more two-dimensional (2D) heat maps associated with key points on line segments corresponding to intersections within the environment; a decoder for decoding the one or more 2D images to determine pixel positions of the key points within the image and semantic labels corresponding to the pixel positions; A path generator, configured to: sample key point pairs based at least on heading directions corresponding to the key point pairs to determine connectable pairs; and generating one or more suggested paths through the intersection by connecting the connectable pairs of key points; as well as A control assembly is configured to perform one or more operations by the vehicle using the one or more suggested paths.

18. The system of claim 17, wherein the 2D heat map comprises an unnormalized normal distribution, and wherein the 2D heat map comprises one of a fixed standard deviation or an adaptive covariance.

19. The system of claim 17, wherein sampling the keypoint pairs further comprises: include: Based at least on analyzing traffic regulations corresponding to the area where the intersection is located.

20. The system of claim 17, wherein each of the 2D heatmaps corresponds to a single semantic label, and the semantic label for each keypoint is determined based at least on the 2D heatmap from which the keypoint was decoded.