Landmark detection using curve fitting for autonomous driving applications

Abstract

In various examples, one or more deep neural networks (DNNs) are executed to regress control points of curves, and the control points can be used to perform a curve fitting operation (e.g., a Bezier curve fitting) to identify landmark locations and geometry in an environment. Thus, the output of the DNN(s) can indicate two-dimensional (2D) image space and / or three-dimensional (3D) world space control point locations, and post-processing techniques such as clustering and temporal smoothing can be performed to accurately and in real-time determine landmark locations and poses. Thus, a vehicle can use reconstructed curves corresponding to landmarks (e.g., lane lines, road boundary lines, crosswalks, poles, text, etc.) to perform one or more operations for navigating an environment.

Description

BACKGROUND

[0001] The ability to accurately and precisely detect landmarks (e.g., lane lines, lane edges, road boundaries, text, poles or signs, and / or other features) in an environment is necessary for all levels of autonomous machine applications (e.g., from driver / pilot-assisted machines or vehicles to semi-autonomous and fully autonomous machines and vehicles). The variety of lane markings that can be encountered in an environment during driving can be very high due to variations in road marking quality, geographic or regional differences in lane and road marking conventions, and due to road marking obstructions, degradation, and / or occlusions caused by wear, weather conditions, lighting conditions, temporary markings (e.g., due to construction or disaster relief), etc.

[0002] Some conventional lane or lane line detection methods have used deep neural network (DNN) processing in which a DNN (e.g., a convolutional neural network (CNN)) is trained using high resolution images of a driving surface and associated annotations of lanes and lines to recognize lane lines. These conventional methods have trained the DNN to generate a segmentation mask showing the general location of lane lines in an image by classifying each pixel of the image as either or not as part of a lane line. However, these conventional methods require significant post-processing (such as dynamic programming) to convert the output into a usable representation of lane lines, road boundaries, poles, and / or other features. For example, because the output of these conventional DNNs includes an identification of each of the pixels belonging to each landmark feature, the correlation between each of the pixels belonging to the same landmark and its classification must be determined during post-processing. Furthermore, for landmarks having a profile other than a straight profile (e.g., to follow a curve or profile line of a driving surface), additional angular or directional information must be output by the DNN and post-processed to determine the angle or direction of the landmark feature. These significant post-processing tasks increase the latency of the system and increase the computational requirements of the system, making these conventional DNNs less useful for real-time deployment in vehicles. SUMMARY

[0003] Embodiments of the present disclosure relate to landmark detection using curve fitting for autonomous driving applications. Systems and methods are disclosed that use one or more deep neural networks (DNNs) to regress control points of a curve and perform a curve fitting operation (e.g., a Bezier curve fitting) using the control points to identify landmark locations and geometry in an environment. Thus, the output of the DNN(s) can indicate two-dimensional (2D) image space and / or three-dimensional (3D) world space control point locations, and less computationally and time intensive post-processing techniques such as clustering and temporal smoothing can be performed to accurately and in real-time determine landmark locations and poses. Thus, in comparison to conventional systems, for each landmark in the field of view of each of one or more sensors of a vehicle, control points can be regressed similar to object detection methods, and then the control points can be used to reconstruct a curve corresponding to the landmark (e.g., a lane line, a road boundary line, a crosswalk, a pole, text, etc.).

[0004] To train the DNN(s) for accurate predictions, ground truth data corresponding to control point locations of landmark features can be generated from multi-segment line and / or polygon annotations corresponding to the landmark features in a training dataset. For example, because a training dataset can include previously annotated multi-segment lines or polygons, curve fitting methods can be used to analyze these existing annotations (in addition to or in lieu of new or additional training datasets) to determine control points for generating ground truth data. In some embodiments, to improve the accuracy of the control point determinations, points from the ground truth multi-segment lines can be resampled to create more evenly spaced points along the multi-segment lines so that the determined control points more accurately reflect the curve of the landmark feature. The ground truth data used to train the DNN(s) to regress control points can include a binary mask indicating pixels within a threshold distance from the curve, distance regression channels indicating distances in the x, y, and / or z directions (for 3D control point regression) to the control points, and / or classification channels for determining classification or semantic information corresponding to each control point or fitted curve. Once trained and deployed, the DNN(s) can accurately and precisely compute outputs indicating control point locations and semantic information corresponding thereto, and these outputs can be decoded and post-processed to determine curves corresponding to landmark features. BRIEF DESCRIPTION OF DRAWINGS

[0005] The present systems and methods for landmark detection using curve fitting for autonomous driving applications are described in detail below with reference to the attached drawing figures, wherein:

[0006] Figure 1 is a dataflow diagram illustrating an example process for training a deep neural network (DNN) for landmark detection, in accordance with some embodiments of the present disclosure;

[0007] Figure 2A depicts a visualization of annotations associated with training images according to some embodiments of the present disclosure;

[0008] Figure 2B depicts a curve fitting operation according to some embodiments of the present disclosure;

[0009] Figure 2C depicts a visualization of annotations converted to curves using a curve fitting operation according to some embodiments of the present disclosure;

[0010] Figure 3A depicts a visualization of training data used to generate ground truth data according to some embodiments of the present disclosure;

[0011] Figure 3B depicts a visualization of a binary mask according to some embodiments of the present disclosure;

[0012] Figure 3C depicts a visualization of a bit classification mask according to some embodiments of the present disclosure;

[0013] Figures 3D-3K depicts a visualization of a distance regression mask according to some embodiments of the present disclosure;

[0014] Figures 4A-4C depicts a visualization of ground truth data generated using reduced length and varying weights according to some embodiments of the present disclosure;

[0015] Figure 5 is a flowchart showing a method for training a DNN for landmark detection according to some embodiments of the present disclosure;

[0016] Figure 6 is a dataflow diagram showing an example process using a DNN for landmark detection according to some embodiments of the present disclosure;

[0017] Figure 7 is a dataflow diagram showing an example process of decoding and post-processing the output of a DNN for landmark detection according to some embodiments of the present disclosure;

[0018] Figures 8A-8B depicts a visualization of predicted curves before and after post-processing according to some embodiments of the present disclosure;

[0019] Figures 9A-9D depicts a visualization of example landmark types detected using a DNN according to some embodiments of the present disclosure;

[0020] Figure 10 is a flowchart showing a method for landmark detection using a DNN according to some embodiments of the present disclosure;

[0021] Figure 11A is an illustration of an example autonomous vehicle in accordance with some embodiments of the present disclosure;

[0022] Figure 11B is an illustration of an example autonomous vehicle in accordance with some embodiments of the present disclosure; Figure 11A is an illustration of camera positions and fields of view of an example autonomous vehicle;

[0023] Figure 11C is an illustration of an example autonomous vehicle in accordance with some embodiments of the present disclosure; Figure 11A is a block diagram of an example system architecture of an example autonomous vehicle;

[0024] Figure 11D is a system diagram for communicating between a cloud-based server and an example autonomous vehicle in accordance with some embodiments of the present disclosure; Figure 11A is a system diagram for communicating between a cloud-based server and an example autonomous vehicle in accordance with some embodiments of the present disclosure;

[0025] Figure 12 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure; and

[0026] Figure 13 is a block diagram of an example data center suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0027] Systems and methods related to landmark detection using curve fitting for autonomous driving applications are disclosed. Although the present disclosure can be described with respect to an example autonomous vehicle 1100 (alternatively referred to herein as “vehicle 1100” or “autonomous vehicle 1100,” examples of which are described herein with respect to Figures 11A-11D , this is not intended to be limiting. For example, the systems and methods described herein can be used by a non-autonomous vehicle, a semi-autonomous vehicle (e.g., in an adaptive driver assistance system (ADAS)), a robot, a warehouse vehicle, an off-road vehicle, a dirigible, a ship, a passenger vehicle, a car, a truck, a bus, a first responder vehicle, a shuttle, an electric or motorized bicycle, a motorcycle, a fire vehicle, a police vehicle, an ambulance, a construction vehicle, an underwater vehicle, a drone, another type of vehicle (e.g., a vehicle that is unmanned and / or that accommodates one or more passengers), a manned or unmanned robot, and / or a manned or unmanned robotic platform. Moreover, although the present disclosure can be described with respect to autonomous driving, this is not intended to be limiting. For example, the systems and methods described herein can be used in robotics, aerial systems, maritime systems, and / or other technical fields, e.g., for perception, world model management, path planning, obstacle avoidance, and / or other processes.

[0028] See Figure 1 , Figure 1This is a data flow diagram illustrating an example process for training a deep neural network (DNN) for landmark detection according to some embodiments of this disclosure. It should be understood that such and other arrangements described herein are illustrated by way of example only. Other arrangements and elements (e.g., machines, interfaces, functions, sequences, functional groups, etc.) may be used in addition to or in place of those shown, and some elements may be omitted together. Furthermore, many of the elements described herein are functional entities that can be implemented as discrete or distributed components, or in combination with other components and implemented in any suitable combination and location. The different functions described herein as being performed by entities may be performed by hardware, firmware, and / or software. For example, different functions may be performed by a processor executing instructions stored in memory. In some embodiments, at least in part, the arrangements described herein may be used. Figures 11A-11D Vehicle 1100 Figure 12 Example computing device 1200, and / or Figure 13 The example data center 1300 describes similar components, features, and / or functions to implement the training of the deep neural network (DNN) 104 according to process 100.

[0029] Process 100 may include generating and / or receiving sensor data 102 from one or more sensors. As a non-limiting example, this data may originate from a vehicle (e.g., as described herein). Figures 11A-11D The vehicle 1100 receives sensor data 102 from one or more sensors. Sensor data 102 can be received by the vehicle 1100 and used within process 100 to train one or more DNNs 104 to detect landmarks—e.g., landmark location, pose, semantic class, etc. During training, sensor data 102 can be generated using one or more data collection vehicles, and / or can be pre-generated and included in the training dataset, said one or more data collection vehicles generating sensor data for training the DNNs (such as, (one or more) DNNs 104). Sensor data 102 used during training can additionally or alternatively be generated using simulated sensor data (e.g., sensor data generated using one or more virtual sensors of a virtual vehicle in a virtual environment) and / or enhanced sensor data (e.g., sensor data generated using one or more data collection vehicles and modified using virtual data). Once trained and deployed in vehicle 1100, sensor data 102 can be generated by one or more sensors of vehicle 1100 and processed by (one or more) DNNs 104 to compute different outputs 106, as described herein.

[0030] Thus, sensor data 102 may include, but is not limited to, sensor data 102 from any sensor of vehicle 1100, for example and reference.Figures 11A-11C Any sensors of vehicle 1100 include RADAR sensor(s) 1160, ultrasonic sensor(s) 1162, LIDAR sensor(s) 1164, stereo camera(s) 1168, wide-view camera(s) 1170 (e.g., fisheye cameras), infrared camera(s) 1172, surround camera(s) 1174 (e.g., 360-degree cameras), long and / or mid-range camera(s), and / or other sensor types. As another example, sensor data 102 can include virtual (e.g., simulated or augmented) sensor data generated from any number of sensors of a virtual vehicle or other virtual object in a virtual (e.g., test) environment. In such examples, the virtual sensors can correspond to a virtual vehicle or other virtual object in a simulated environment (e.g., for testing, training, and / or validating DNN performance), and the virtual sensor data can represent sensor data captured by the virtual sensors within the simulated or virtual environment. Thus, by using virtual sensor data, the DNN(s) 104 described herein can be tested, trained, and / or validated using simulated or augmented data in a simulated environment, which can allow for testing of more extreme scenarios outside of real-world environments where such testing can be less safe.

[0031] In some embodiments, sensor data 102 can include image data representing image(s), image data representing video (e.g., a snapshot of a video), and / or sensor data representing a representation of a sensor’s sensory field (e.g., a depth map of a LIDAR sensor, a value map of an ultrasonic sensor, etc.). Where sensor data 102 includes image data, any type of image data format can be used, such as but not limited to compressed images like Joint Photographic Experts Group (JPEG) or luminance / chrominance (YUV) formats, as compressed images sourced from frames of a compressed video format like H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC), raw images sourced from a Red Clear Blue (RCCB), Red Clear Clear (RCCC), or other type of imaging sensor, and / or other formats. Additionally, in some examples, sensor data 102 can be used within process 100 without any pre-processing (e.g., in raw or captured format), while in other examples, sensor data 102 can undergo pre-processing (e.g., noise balancing, demosaicing, scaling, cropping, enhancement, white balancing, tone curve adjustment, etc., such as using a sensor data pre-processor (not shown)). As used herein, sensor data 102 can refer to unprocessed sensor data, pre-processed sensor data, or a combination thereof.

[0032] The sensor data 102 used for training can include raw images (e.g., as captured by one or more image sensors), downsampled images, upsampled images, cropped or region of interest (ROI) images, otherwise augmented images, and / or combinations thereof. The DNN(s) 104 can be trained using the images (and / or other sensor data 102) along with corresponding ground truth data 122 (e.g., ground truth (GT) binary mask(s) 124, GT distance regression mask(s) 126, and / or GT bit classification mask(s) 128). The ground truth data 122 can be generated by the encoder 120 using annotation data 114 (e.g., annotations, labels, masks, etc.), curve fitters 116, and / or control points 118. For example, in some embodiments, the annotation data 114 can represent a polyline corresponding to any landmark type (e.g., road marking, lane line, road boundary line, crosswalk line, sidewalk, bike lane line, text, pole, tree, light pole, sign, etc.). For example, with respect to Figure 2A The visualization 200A can include polylines 202A-202F, which can correspond to lane and / or road boundary lines. The polylines 202A-202F can include corresponding semantic or classification information. For example, when the polylines 202 are annotated, the polylines 202 can be associated with a particular semantic class. For example, the polyline 202A can correspond to a road boundary line, or more finely, a solid white road boundary line. Similarly, the polyline 202C can correspond to a dashed lane line, and the polyline 202D can correspond to a solid road boundary line.

[0033] In some embodiments, a polyline can be determined or extracted from a polygon. For example, an annotation or label represented by the annotation data 114 can correspond to a polygon that substantially defines a shape of a landmark - e.g., a polygon that fits to each lane, a polygon that fits to a pole, a polygon that fits to text, etc. A centerline, left boundary line, and / or right boundary line of the polygon can be used to produce a polyline. For example, for a lane, a polygon can be used to determine a left lane line (e.g., a left edge of the polygon), a right lane line (e.g., a right edge of the polygon), and / or a centerline or track (e.g., a line that extends through the center of the polygon) of the lane. Thus, in cases where a polygon is used, a polyline can be produced or extracted from the polygon, e.g., the polyline 202 of Figure 2A In any embodiment, the annotation can include a polygon, a polyline, another annotation or label type, or combinations thereof.

[0034] As described herein, the annotation data 114 can represent semantic or classification information corresponding to polygons, polylines, and / or other annotation or label types. The semantic categories can include any level of granularity. For example, the semantic categories can include lane lines, road boundary lines, crosswalk lines, poles, signs, text, etc. At a more granular level, for example, the semantic categories for lane lines can include solid or dashed, yellow or white, single or double, or combinations thereof. Thus, for non-limiting examples, the semantic categories for lane lines can include white single solid, yellow single solid, white single dashed, yellow single dashed, white double solid, yellow double solid, white double dashed, yellow double dashed, white solid-dashed, yellow solid-dashed, white dashed-solid, yellow dashed-solid, triple solid, triple dashed, solid-dashed-solid, dashed-solid-dashed, etc. For road boundaries, for non-limiting examples, the semantic categories can include obstacles, curbs, joints, lines, etc. For crosswalk lines, for non-limiting examples, the semantic categories can include crosswalk lines, x-crosswalk lines, stop lines, pedestrian lines, etc. For poles, for non-limiting examples, the semantic categories can include road signs, poles, lights, utility poles, trees, etc., or for signs can include traffic signs, construction signs, advertising signs, etc. Thus, any number and granularity of semantic categories can be associated with annotations or labels and used to generate ground truth data 122 for training the DNN 104.

[0035] In some examples, the annotations can be generated within a drawing program (e.g., annotation program), a computer aided design (CAD) program, a labeling program, another type of program adapted to generate annotations, and / or the annotations can be manually drawn. In any example, the annotation data 114 can be synthetically produced (e.g., generated from a computer model or rendering), authentically produced (e.g., designed and produced from real world data), machine automated (e.g., using feature analysis and learning to extract features from data and then produce labels), human annotated (e.g., a labeler or annotation expert defines the location of a label), and / or combinations thereof (e.g., a human identifies the center or origin and dimensions of a region, a machine generates a polygon and / or label for the intersection region).

[0036] The process 100 can include performing a curve fitting operation using the curve fitter 116. For example, the polyline represented by the annotation data 114 can undergo a curve fitting operation to determine control points that bound the points of the curve. Thus, the ground truth data 122 can be used to train the DNN(s) 104 to predict, regress, or otherwise calculate the locations of the control points in 2D image space or 3D world space. The curve fitting operation can include a 2D Bezier curve fitting algorithm (e.g., for image space control point estimation), a 3D Bezier curve fitting algorithm (e.g., for world space control point estimation), a cubic Bezier curve fitting algorithm (e.g., containing four control points for each curve), a higher order Bezier curve fitting algorithm (e.g., five or more control points), or a split Bezier curve fitting algorithm (e.g., a single curve is split into two or more curves, each containing multiple control points), and / or another curve fitting algorithm type. In some embodiments, the curve fitting algorithm can be used to determine three or more control points in order to improve the precision and accuracy of the curve fitting. For example, with only two points, the result of the curve fitting is similar to the result of a linear interpolation, which would produce less accurate or precise results for a non-linear curve. Thus, the more control points that are used, the more accurate or precise the curve fitting, while the more control points, the more processing and more computationally intensive the operation, thereby increasing the run time of the system. Thus, through experimentation, and by way of non-limiting example, using three to five control points can produce accurate and precise results with acceptable computational requirements that allow for real-time deployment. However, any number of control points can be used without departing from the scope of the present disclosure. Furthermore, while the illustrations herein include four control points per curve, this is not intended to be limiting and is for example purposes only. Further, while Bezier curve fitting is primarily described herein, this is not intended to be limiting and other curve fitting algorithms can be used without departing from the scope of the present disclosure. However, Bezier curve fitting includes several advantages, including the use of three or more control points to generate more accurate curve fitting results, and two of these control points are located at the start and end point of each curve, which is useful for the vehicle 1100 in making plans, controls, or other decision types.

[0037] In some embodiments, prior to the curve fitting operation, the points along the polyline from the annotation data can be resampled. For example, because the polyline can contain fewer points where the curve is straight and more points where there is more curvature, there can be more points at some locations than others. While a polyline that is not resampled can produce accurate results, the accuracy can be increased by resampling the points along the polyline to contain a more evenly sampled distribution of points. By way of example, Figure 2AThe polyline 202F can include points 206 prior to resampling, while the polyline 202C can include points 206 after resampling. As illustrated, the points 206 along the polyline 202C are more evenly distributed than the points along the polyline 202F. With resampling, the curve fitting operation can be less likely to produce control points that are biased toward areas of the polyline with a greater number of points. Thus, the control points determined using the curve fitter 116, and thus the ground truth data 122 generated using the control points, can be more accurate for curve fitting.

[0038] In embodiments, after resampling, the polyline can then undergo a curve fitting operation to determine control point locations. Referring to Figure 2B And as an example using a Bezier curve fit, a Bezier curve can be defined as Q(t), where t can include a parameter such that 0 < t < 1. The following in a matrix representation of equation (1), a first matrix T is composed of t 0 , t 1 , t 2 , and t 3 In equation (1), a second matrix M can be referred to as a Bezier matrix, which is composed of the coefficients when we expand the cubic polynomial. The four control points (in this example, a cubic Bezier curve fit algorithm) include P0, P1, P2, and P3 and, since there are two dimensions in this example, the control points can be represented in equation (1) as a third matrix P, where x0 and y0 represent the (x, y) coordinates (e.g., pixel coordinates in a 2D image space) of control point P0, and so on. Although represented as 2D pixel or image space coordinates, this is not intended to be limiting, and in some embodiments, the coordinates can correspond to 3D world space coordinates, such that the DNN 104 is trained to directly compute the 3D coordinates of the control points of the curve.

[0039]

[0040] The Bezier curve fit algorithm can alternately optimize P and T. An initial T can be computed by the length of each line segment (or polyline). If there is only one line segment, the line segment can be equally divided into three line segments. If there are only two line segments, the longer line segment can be equally divided into two line segments. In embodiments, any line segment can be divided into small pieces such that each line segment has a relatively same length. The initial T can be computed after the line segments have been determined, and then can be optimized using the resampled polyline. When T is fixed, P can be solved by least squares, such as according to the following equation (2):

[0041] P = (M T T T TM) -1 M T TT Q (2)

[0042] When P is fixed, T can be solved by Newton-Raphson method. For example, the problem to be solved can be the root of the following equation (3):

[0043] f(t) = (Q(t) - Q)Q'(t) (3)

[0044] As an example, the concept is illustrated in Figure 2B where f(t), f'(t) are computed according to the following equation (4):

[0045] f'(t) = Q'(t)Q'(t) + (Q(t) - Q)Q"(t) (4)

[0046] where the update step is represented by the following equation (5):

[0047] t = t - f(t) / f'(t) (5)

[0048] At each iteration, the error e can be computed according to the following equation (6):

[0049] e = ||P - TMQ|| (6)

[0050] If the error e stops decreasing, convergence can be determined. If the error e exceeds a maximum error, divergence can be determined. If a maximum number of steps or iterations is reached, the result of the last iteration can be used.

[0051] As a result of the curve fitting operation of the curve fitter 116, one or more curves defined by the control points 118 can be generated. As an example of a result of the curve fitting operation, Figure 2C a visualization 200B of curves 204 (e.g., curves 204A through 204F) generated from the polyline 202 of Figure 2A after curve fitting. As such, each curve (of a particular semantic label) includes an equal number of control points 118 - e.g., four in this example, as the case where a Bezier curve fitting is performed three times. As shown, even though continuous in the world space, two curves 204 can be generated each having respective control points 118 because curve 204C corresponds to a lane line and curve 204E corresponds to a road boundary line. As a result, as described herein, the DNN(s) 104 can be trained to compute the bit classification mask(s) 112 corresponding to a particular semantic class, and / or to regress (or otherwise compute) the binary mask(s) 108 and distance regression mask(s) 110 of the control points 118 of the curve 204 corresponding to a particular landmark type (e.g., a particular road marking type with respect to Figure 2C ​

[0052] The control points 118 and / or semantic classification information can then be used by the encoder 120 to generate ground truth data 122 for training the DNN(s) 104 using the training engine 130. As such, for each instance of sensor data 102 in the training dataset, ground truth data 122 can be generated. The ground truth data 122 can include one or more GT binary masks 124, one or more GT distance regression masks 126, and / or one or more GT bit classification masks 128. In some embodiments, a weight channel can also be generated for the ground truth data 122, as in the case where variable length and / or width are used to encode the curves in the ground truth data 122, as described in greater detail herein. Ultimately, the training engine 130 can compare the ground truth data 122 to the output 106 of the DNN(s) 104 using one or more loss functions, such as but not limited to those described herein. For example, the binary mask(s) 108 computed by the DNN(s) 104 for an instance of sensor data 102 can be compared to the GT binary mask(s) 124 corresponding to the same instance of sensor data 102, the distance regression mask(s) 110 computed by the DNN(s) 104 for an instance of sensor data 102 can be compared to the GT distance regression mask(s) 126 corresponding to the same instance of sensor data 102, and so on. The training engine 130 can use the loss function(s) to determine updates to the parameters (e.g., weights, biases, etc.) of the DNN(s) 104, and this process can be repeated until the DNN(s) 104 converge to an acceptable level of accuracy, e.g., until the trained parameters of the DNN(s) 104 converge to optimal, desired, and / or acceptable values.

[0053] As an example, and with reference to Figure 3AThe visualization 300 may correspond to instances of sensor data 102 (e.g., images) and corresponding curves 302A-302D, each curve including control points 118 (e.g., each curve 302 may include four control points 118). Each curve 302 may have been generated by curve fitter 116 from a polyline represented by annotation data 114. Each curve 302 may be associated with semantic classification information. For example, curves 302A and 302D may correspond to road boundary lines, and curves 302B and 302C may correspond to lane lines (e.g., lane edge lines). As illustrated only with respect to curve 302D (for clarity), each curve 302 may include a first control point 118A, a second control point 118B, a third control point 118C, and a fourth control point 118D. For each curve, the first control point 118A may correspond to an end control point, while the fourth control point 118D may correspond to a start control point 118.

[0054] Encoder 120 can use curve 302 to encode GT binary mask 124. For example, regarding Figure 3B The diagram illustrates a GT binary mask 124A corresponding to curve 302 from visualization 300. The GT binary mask 124A can be generated by encoding a first value (e.g., 1) for each pixel within a predefined radius of any curve 302 and a second value (e.g., 0) for each pixel outside the predefined radius of any curve 302. The radius can be chosen as any value. For example, in a non-limiting example where the input and output resolutions of DNN 104 are different (e.g., due to downsampling via convolution operations), the radius can be between 2 and 6 pixels at the output resolution of DNN 104, and can include 16 to 48 pixels at the input resolution of DNN 104 (e.g., the resolution of the image from visualization 300) (e.g., where the resolution difference is 8x). However, different radii can be set depending on the embodiment and / or the difference between the input and output resolutions can be set depending on the embodiment. For example, since pixels encoded as being within the radius can be used as candidate pixels for regressing (or otherwise calculating) control points 118 during decoding, the radius can be determined such that enough candidate pixels are selected for accurate and precise predictions, without including so many pixels that would increase processing requirements and runtime.

[0055] The encoder 120 can encode the GT bit classification mask 128 using the control points 118 of the curves 302. For example, the GT bit classification mask 128 or the number of channels can be determined based on the number of different semantic classes that the DNN 104 is trained to predict. For example, if there are n semantic classes, there can be an upper limit (maximum) of log2(n) channels. Thus, if there is only a single semantic class, there can be no bit classification channels, and thus no GT bit classification mask 128. However, where there are two bit classification channels, there can be a single bit classification channel and thus a single GT bit classification mask 128. In this example, a first value (e.g., 0) can indicate that a pixel corresponds to a first semantic class, and a second value (e.g., 1) can indicate that a pixel corresponds to a second semantic class. As an example, and with respect to the visualization 300, the GT bit classification mask 128A can correspond to the visualization 300 where there are two semantic classes (e.g., assume that the DNN 104 is only trained to predict for two classes in this example). For example, the curves 302C and 302B corresponding to lane lines can be encoded with a first value (e.g., 1, corresponding to white), and the curves 302A and 302D corresponding to road boundary lines can be encoded with a second value (e.g., 0, corresponding to black). In some embodiments, each of the pixels of the GT distance regression mask 126 can be encoded (or considered for encoding) with a bit classification value, while in other examples, only the pixels of the GT binary mask 124 that are identified as being within the radius of the curves 302 can be encoded with a bit classification value. Where a radius is used for each mask type, the DNN 104 can be trained to predict pixels that are most likely to contribute to the accuracy of the prediction, and not background pixels. Figure 3C

[0056] ​The encoder 120 can encode the GT distance regression masks 126 using the control points 118 of the curves 302. For example, for each control point 118, the x-axis GT distance regression mask 126 and the y-axis GT distance regression mask 126 can be encoded. In embodiments where the control points 118 are in 3D world space coordinates, each control point 118 can include an additional z-axis GT distance regression mask 126. For example, each curve 302 can include a first control point 118 (e.g., an ending control point), the pixels of the x-axis GT distance regression mask 126 can be encoded with a distance (e.g., a pixel distance) to the nearest first control point 118A of any of the curves 302 in the x-direction, and the pixels of the y-axis GT distance regression mask 126 can be encoded with a distance (e.g., a pixel distance) to the nearest first control point 118A of any of the curves 302 in the y-direction. This process can be repeated for each of the control points 118B-118D (where there are four control points 118, but can include more or fewer iterations where there are more or fewer control points 118). The values of the distances can have a scaling factor applied to them to produce updated distance values that are within a smaller range of values (e.g., closer to 0-1 than 0-90 or some other values). In some embodiments, each of the pixels of the GT distance regression masks 126 can be encoded, while in other examples, only the pixels of the GT binary masks 124 that are identified as being within the radius of the curves 302 can be encoded with the distance values. In examples where only the pixels within the radius are used, the accuracy can not be affected while the processing requirements and run time can be reduced.

[0057] As an example, and with respect to Figures 3D to 3K , the control point 118A of each curve 302A-302D can be encoded into an x-axis GT distance regression mask 126A-X( Figure 3D ) and a y-axis GT distance regression mask 126A-Y( Figure 3E ), the control point 118B of each curve 302A-302D can be encoded into an x-axis GT distance regression mask 126B-X( Figure 3F ) and a y-axis GT distance regression mask 126B-Y( Figure 3G ), the control point 118C of each curve 302A-302D can be encoded into an x-axis GT distance regression mask 126C-X( Figure 3H ) and a y-axis GT distance regression mask 126C-Y( Figure 3I ), and the control point 118D of each curve 302A-302D can be encoded into an x-axis GT distance regression mask 126D-X( Figure 3J ) and a y-axis GT distance regression mask 126D-Y( Figure 3K ).

[0058] In some embodiments, the distance can correspond to an output resolution of the DNN 104, while in other embodiments, the distance can correspond to an input resolution of the DNN 104. Thus, even though the x-axis or y-axis distance can be computed for a pixel at the output resolution, the value can correspond to a distance (e.g., a pixel distance) at the input resolution (e.g., of the input image of the DNN 104). Thus, during decoding, a pixel at the output resolution can have a known or estimated corresponding pixel at the input resolution, and the known or estimated corresponding pixel can then have a value from the output resolution attributed to it. This process can be similar to that in U.S. Non-Provisional Patent Application No. 16 / 514,230, filed July 17, 2019, entitled Figures 2A-2B and the accompanying description, which is hereby incorporated by reference in its entirety.

[0059] In some embodiments, such as where two or more of the GT binary mask 124, GT distance regression mask 126, or GT bit classification mask 128 are generated so that only pixels within a radius of the curve 302 are encoded, the radius can be different for different mask types. For example, the GT binary mask 124 can have a first radius, and the GT distance regression mask 126 can have a second radius that is different than the first radius. Similarly, the GT bit classification mask 128 can have a different radius than the GT binary mask 124 and / or the GT distance regression mask 126. In such embodiments, another GT weight channel or mask can be generated to train the DNN 104 to output a weight channel to help decode the masks, even though their radii (e.g., the radius of pixels from the curve) can be different. For example, with respect to Figures 4A-4C , the GT binary mask 124B( Figure 4A ) can include a first radius, the GT distance regression mask 126E( Figure 4B ) can include a second, larger radius, and the weight channel or mask 400( Figure 4C ) can be used to encode this information so that the decoder 120( Figure 6 ) can use the weight channel computed by the DNN 104 during deployment to more accurately regress (or otherwise compute) the locations of the control points 118.

[0060] In some embodiments, to make the predictions of the DNN(s) 104 more focused, the ground truth data 122 can be encoded with shorter curve lengths. For example, each of the GT binary mask(s) 124, GT distance regression mask(s) 126, and / or GT bit classification mask(s) 128 can be encoded with shorter distances - as by Figures 3B to 3K the mask of Figures 4A to 4Cdifferences between the masks. For example, in some embodiments, the encoded values for the curves in each mask can be cropped at the top and / or bottom by some percentage or amount. In non-limiting embodiments, values corresponding to the top 25% of the curve and values corresponding to the bottom or lower 25% of the curve can be cropped, such that the resulting mask from the ground truth data 122 is 50% of its potential or original length (the entire curve would be encoded). These values are for example purposes only, and one or more key performance indicators (KPIs) can be used to determine suitable or ideal values for any particular embodiment. This process can be helpful because it takes into account the receptive field of the DNN 104. For example, if a full curve or polyline is used, the pixels or points on each end of the curve or polyline (e.g., in 3D world space) can have the same accuracy as the regression channel output.

[0061] Accordingly, once the ground truth data 122 has been encoded, the DNN(s) 104 can undergo training using the training engine 130. The sensor data 102 can be applied to the DNN(s) 104, and the DNN(s) 104 can compute the output 106. The output 106 can then be compared to the corresponding ground truth data 122 using one or more loss functions, and the training engine 130 can use the output of the loss function(s) to update the parameters of the DNN(s) 104. This process can be repeated until acceptable convergence has been achieved.

[0062] In some embodiments, the loss function can be different for different ground truth data 122 types. For example, a first loss function can be used to train the DNN 104 for the GT binary mask 124 and the binary mask 108, a second loss function can be used to train the DNN 104 for the GT distance regression mask 126 and the distance regression mask 110, and a third loss function can be used to train the DNN 104 for the GT bit classification mask 128 and the bit classification mask 112. As non-limiting examples, a binary cross-entropy loss can be used to train the binary mask channel, a mask LI loss can be used to train the distance regression channel, and a mask binary cross-entropy loss can be used to train the bit classification channel.

[0063] Referring now to Figure 5Each block of the method 500 described herein comprises a computational process that can be performed using any combination of hardware, firmware, and / or software. For instance, various functions can be carried out by a processor executing stored instructions. The method 500 can also be embodied as computer-usable instructions stored on computer storage media. The method 500 can be provided by a standalone application, a service or hosted service, an add-on to another product, or a plug-in to another product, to name a few. Figure 1 The method 500 is described with respect to the process 100 of FIG. 1 by way of example. However, this method 500 can additionally or alternatively be performed by any one process and in any one system or any combination of processes and systems, including but not limited to those described herein.

[0064] Figure 5 FIG. 5 is a flowchart illustrating a method 500 for training a DNN 104 for landmark detection, in accordance with some embodiments of the present disclosure. At block B502, the method 500 includes receiving annotation data representing a polyline associated with an image and classification information associated with the polyline. For example, the annotation data 114 associated with an instance of sensor data 102 can be received and / or generated, and the annotation data can represent one or more polylines corresponding to a landmark and / or classification information associated therewith.

[0065] At block B504, the method 500 includes performing a curve fitting algorithm based at least in part on the polyline to determine a plurality of control points associated with a curve. For example, the curve fitter 116 can perform a curve fitting operation using the annotation data 114 to determine control points corresponding to a curve.

[0066] At block B506, the method 500 includes encoding a binary mask indicating pixels of the image within a radius of the curve and encoding a plurality of distance regression masks based at least in part on the plurality of control points associated with the curve. For example, using the control points 118 and / or curve information, the encoder 120 can encode the GT binary mask 124 and / or the GT distance regression masks 126, where each distance regression mask can indicate a distance from a control point 118 of the plurality of control points 118.

[0067] At block B508, the method 500 includes encoding one or more bit classification masks based at least in part on the classification information. For example, the encoder 120 can encode the GT bit classification mask 128 using the classification information and / or control point or curve information.

[0068] At block B510, the method 500 includes training the neural network using the binary mask, the plurality of distance regression masks, and the one or more bit classification masks as ground truth data. For example, the training engine 130 can use the ground truth data 122 to train the DNN 104 to compute the output 106.

[0069] DNN for landmark detection

[0070] Referring now to Figure 6 , Figure 6 is a dataflow diagram illustrating an example process 600 of using a DNN for landmark detection, in accordance with some embodiments of the present disclosure. It should be understood that this and other arrangements described herein are set forth only as examples. Other arrangements and elements (e.g., machines, interfaces, functions, orders, groupings of functions, etc.) can be used in addition to or instead of those shown, and some elements can be omitted altogether. Further, many of the elements described herein are functional entities that can be implemented as discrete or distributed components or in conjunction with other components, and in any suitable combination and location of components, even though only one of each such element is described for clarity. The different functions described herein as being performed by various entities can be performed by hardware, firmware, and / or software. For instance, different functions can be performed by a processor executing stored instructions. In some embodiments, the deployment of the DNN(s) 104 according to the process 600 can be implemented at least in part using similar components, features, and / or functionality as described herein with respect to the vehicle 1100 of Figures 11A-11D the example computing device 1200 of Figure 12 and / or the example data center 1300 of Figure 13 .

[0071] The process 600 can include generating and / or receiving sensor data 102 generated by one or more sensors of the vehicle 1100. The sensor data 102 can be similar to the sensor data described herein, as described with respect to the process 100 of Figure 1 The sensor data 102 can be applied to the DNN(s) 104, which can be trained according to the process 100 of Figure 1The process 100 is trained - and the DNN(s) 104 can compute the output 106. In embodiments, the DNN(s) 104 can include a convolutional neural network (CNN). Although examples are described herein with respect to using a neural network, specifically a convolutional neural network, as the DNN 104, this is not intended to be limiting. For example, and without limitation, the DNN 104 described herein can include any type of machine learning model, such as machine learning models using linear regression, logistic regression, decision trees, support vector machines (SVMs), naive Bayes, k-nearest neighbors (Knn), K-means clustering, random forests, dimensionality reduction algorithms, gradient boosting algorithms, neural networks (e.g., autoencoders, convolutional, recurrent, perceptrons, long / short-term memory (LSTM), hopfield, Boltzmann, deep belief, deconvolutional, generative adversarial, liquid machines, etc.), and / or other types of machine learning models.

[0072] As an example, such as in the case where the DNN 104 includes a CNN, the DNN 104 can include any number of layers. One or more of the layers can include an input layer. The input layer can 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 can 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).

[0073] One or more of the layers can include a convolutional layer. The convolutional layer can compute outputs of neurons connected to local regions in the input layer, each neuron computing a dot product between their weights and the small region in the input volume they are connected to. The result of the convolutional layer can be another volume, with one dimension based on the number of filters applied (e.g., width, height, and number of filters, such as 32x32x12, if 12 is the number of filters).

[0074] One or more of the layers can include a deconvolutional layer (or transposed convolutional layer). For example, the result of the deconvolutional layer can be another volume with higher dimensions than the input dimensions of the data received at the deconvolutional layer.

[0075] One or more of the layers can include a rectified linear unit (ReLU) layer. The ReLU layer can apply an element-wise activation function, such as max(0, x), for example, zero thresholding. The resulting volume of the ReLU layer can be the same volume as the input to the ReLU layer.

[0076] One or more of the layers can include a pooling layer. The pooling layer can perform a down-sampling operation along the spatial dimensions (e.g., height and width), which can result in a smaller volume than the input to the pooling layer (e.g., 16 x 16 x 12 from a 32 x 32 x 12 input volume).

[0077] One or more of the layers can include one or more fully connected layers. Each neuron in a fully connected layer can be connected to every neuron in the previous volume. The fully connected layer can compute class scores, and the resulting volume can be 1 x 1 x number of classes. In some examples, the CNN can include a fully connected layer such that the output of one or more of the layers of the CNN can be provided as input to the fully connected layer of the CNN. In some examples, one or more convolutional streams can be implemented by the DNN 104, and some or all of the convolutional streams can include respective fully connected layers.

[0078] In some non-limiting embodiments, the DNN(s) 104 can include a series of convolutional layers and max-pooling layers to facilitate image feature extraction, followed by multi-scale dilated convolutional layers and up-sampling layers to facilitate global contextual feature extraction.

[0079] Although the input layer, convolutional layers, pooling layers, ReLU layers, and fully connected layers are discussed herein with respect to the DNN 104, this is not intended to be limiting. For example, additional or alternative layers can be used in the DNN 104, such as normalization layers, SoftMax layers, and / or other layer types.

[0080] In embodiments in which the DNN 104 includes a CNN, different orders and numbers of layers of the CNN can be used, depending on the embodiment. In other words, the order and number of layers of the DNN(s) 104 are not limited to any one architecture.

[0081] Further, some layers can include parameters (e.g., weights and / or biases), such as the convolutional layers and the fully connected layers, while other layers can not include parameters, such as the ReLU layers and the pooling layers. In some examples, the parameters can be learned by the DNN 104 during training. Further, some layers can include additional hyperparameters (e.g., learning rate, stride, epoch, etc.), such as the convolutional layers, the fully connected layers, and the pooling layers, while others can not, such as the ReLU layers. The parameters and hyperparameters are not limited and can differ depending on the embodiment.

[0082] The output 106 can be decoded using the decoder 602 to determine the control points 118. For example, each pixel from the binary mask 108 that is classified as positive (or within the radius of the curve) can be considered a candidate pixel (e.g., for 2D prediction) or point (e.g., for 3D prediction). The corresponding values for the same pixel in the distance regression masks 110 and the bit classification masks 112 can be used as attributes of the curve candidates (e.g., Bezier curve candidates). For example, for a pixel of the binary mask that is a positive pixel (e.g., a white pixel or a pixel identified as being within the radius of the curve), the x distance value from one or more (e.g., all) of the x-axis distance regression masks 110 and the y distance value from one or more (e.g., all) of the y-axis distance regression masks 110 can be used to determine a candidate control point position. This process can be repeated for each of the candidate pixels. Finally, for each landmark depicted in the image, there can be many overlapping or similar control points and curves, as shown in Figure 8A For example, for a single landmark - e.g., a lane line - there can be multiple sets of control points 118. The curve fitter 116 can use each set of control points 118 to perform a curve fitting operation to produce a curve (e.g., a Bezier curve).

[0083] Due to the possibility of generating multiple curves for the same landmark, different post-processing operations can be performed using the post-processor 604 to determine the final curves used by the vehicle 1100 - e.g., used by the autonomous driving software stack (or driving stack) 608. For example, clustering can be performed by the post-processor 604 using any clustering algorithm type. In a non-limiting embodiment, a density-based spatial clustering of applications with noise (DBSCAN) algorithm can be performed on the curve candidates, while in another non-limiting embodiment, a probabilistic voting algorithm can be used. In any example, a similarity function can be used within the clustering algorithm. In some embodiments, the similarity function can include an integral loss similarity function. For example, the following equation (7) can represent an example of an integral loss similarity function.

[0084]

[0085] In equation (7), Pi and P2 are points on the two (Bezier) curves being compared, t is a parameter of the curve moving from 0 to 1, ai and a2 are the cubic coefficients of the first and second curves, respectively, bi and b2 are the quadratic coefficients of the first and second curves, respectively, ci and c2 are the linear coefficients of the first and second curves, respectively, di and d2 are the constant term coefficients of the first and second curves, respectively. Thus, Pi(t) = ai*t 3 + bi*t 2 + ci*t + di (0 <= t <= 1) and P2(t) = a2*t 3 + b2*t 2+c2*t+d2 (0 <= t <= 1) can correspond to a formula for points on two Bezier curves, respectively. After clustering, there should be a single curve for each landmark feature. However, in some cases, there can still be two or more curves for one or more landmark features. In such examples, another post-processing technique of duplicate removal can be performed by the post-processor 604. For example, curves that are close to each other (e.g., within a certain threshold distance from each other) can be compared to each other, and curves with lower confidence can be removed. Thus, based on the predictions of the DNN 104 and / or clustering, each curve can have a confidence associated with it, and curves with the lowest confidence can be removed.

[0086] As an example, and with respect to Figure 8A , the decoded output 106 of the DNN(s) 104 can be represented in the visualization 800A. As shown, multiple candidate curves 802 can be generated for each landmark type (e.g., lane lines or road boundary lines in the Figure 8A exhibition). Thus, post-processing can be performed to generate a final curve 804 including one final curve 804 for each landmark feature. For example, candidate curve 802A can undergo post-processing to generate final curve 804A, candidate curve 802B can undergo post-processing to generate final curve 804B, and so on. These final curves 804 can be associated with the landmarks 606, such that the location, orientation, pose, and / or other information about the landmarks can be known.

[0087] Referring to Figure 7 , Figure 7is a dataflow diagram illustrating an example process 700 of decoding and post-processing outputs of the DNN 104 for landmark detection, according to some embodiments of the present disclosure. For example, the process 700 can include an end-to-end (E2E) process for landmark detection. Different categories of landmarks 606 can be used together to determine final information, such as lane detection results, intersection detection results, road marking detection results, and so on. Thus, to determine solid / dashed lines (or lane lines), the outputs 106 can be decoded to produce candidate curves, which can be clustered to generate final curves (e.g., to populate a lane marking output), and in cases where a final curve is close to another final curve, duplicate removal can be performed to produce solid / dashed line (or lane line) detection results. Similar processing can be performed for road boundary lines, and a combination of lane lines and road boundary lines can be used to identify each lane of a travel surface— e.g., to generate a lane graph. As another example, similar processing can be performed for intersections or crosswalks (e.g., sidewalks, walkways, and so on), poles or signs, and / or text or other shapes identified using the DNN 104. These outputs— e.g., intersection detection results using crosswalks, intersection detection results using poles or road signs, and road marking detection results— can be used to determine intersection locations, structures, and poses, and / or to determine road marking information. In this way, even where separate categories of information are computed by the DNN(s) 104, detected curves of one or more categories can be used to determine final outputs (e.g., lane graphs, intersection structures, and so on) that can be used by the vehicle 1100 (as by the driving stack 608).

[0088] Referring again to Figure 1 , the landmarks 606 can include any type of landmark, such as but not limited to road markings, lane lines, road boundary lines, crosswalks, sidewalks, bike lane lines, text, poles, trees, light posts, or signs. As described herein, different categories can include any number of subcategories that the DNN 104 is trained to predict. As examples of landmarks 606, Figure 9A a visualization 900A including curves corresponding to lane lines, Figure 9B a visualization 900B including curves corresponding to road boundary lines, Figure 9C a visualization 900C including curves corresponding to crosswalks, and Figure 9D a visualization including curves corresponding to poles and signs.

[0089] Referring now to Figure 10Each block of the method 1000 described herein comprises a computational process that can be performed using any combination of hardware, firmware, and / or software. For instance, various functions can be carried out by a processor executing instructions stored in memory. The method 1000 can also be embodied as computer-usable instructions stored on computer storage media (e.g., memory). By way of example, the method 1000 can be provided as part of an independent application, service, or hosted service (stand-alone service or in combination with another hosted service) or as a plug-in for another product. Further, the method 1000 is described with respect to the process 600 of FIG. 6 and the vehicle 1100 of FIG. 11 by way of example only. However, the method 1000 can additionally or alternatively be performed by any one process and within any one system or any combination of processes and systems, including but not limited to those described herein. Figure 6 Figures 11A-11D The method 1000 is described with respect to the process 600 of FIG. 6 and the vehicle 1100 of FIG. 11 by way of example only. However, the method 1000 can additionally or alternatively be performed by any one process and within any one system or any combination of processes and systems, including but not limited to those described herein.

[0090] Figure 10 is a flowchart illustrating a method 1000 for landmark detection using a DNN, in accordance with some embodiments of the present disclosure. At block B1002, the method 1000 includes applying, to a neural network, sensor data representing an image of a sensory field of a sensor of a machine. For example, the instance of sensor data 102 can be applied to the DNN(s) 104.

[0091] At block B1004, the method 1000 includes computing, using the neural network and based at least in part on the sensor data, first data indicating pixels of the image within a radius of at least one curve of one or more curves identified in the image and second data indicating distances to control points of the one or more curves. For example, the binary mask(s) 108 and the distance regression mask(s) 110 can be computed by the DNN(s) 104 based at least in part on processing the instance of sensor data 102.

[0092] At block B1006, the method 1000 includes decoding the first data and the second data to determine locations of the control points of the one or more curves. For example, the binary mask(s) 108 and the distance regression mask(s) 110 can be decoded using the decoder 602 to determine the locations of the control points 118 (e.g., in a 2D image space or a 3D world space).

[0093] At block B1008, the method 1000 includes generating the one or more curves using a curve fitting algorithm and based at least in part on the locations of the control points. For example, the curve fitter 116 can use the control points 118 to determine a shape of a curve corresponding to a landmark in the environment.

[0094] ​At block B1010, the method 1000 includes performing, by the machine, one or more operations based at least in part on the one or more curves. For example, the vehicle 1000 can use the curves (and / or semantic information corresponding thereto) to perform one or more operations such as world model management, mapping, planning, control, etc.

[0095] Example autonomous vehicle

[0096] Figure 11A is an illustration of an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. The autonomous vehicle 1100 (alternatively referred to herein as “vehicle 1100”) can include, but is not limited to, a passenger vehicle such as a car, truck, bus, first responder vehicle, shuttle, electric or motorized bicycle, motorcycle, fire vehicle, police vehicle, ambulance, boat, construction vehicle, underwater vessel, drone, and / or another type of vehicle (e.g., a vehicle that is unmanned and / or that accommodates one or more passengers). Autonomous vehicles are often described in terms of levels of automation as defined by a department of the United States Department of Transportation, the National Highway Traffic Safety Administration (NHTSA), and the Society of Automotive Engineers (SAE) “Taxonomy and Definitions for Terms Related to Driving Automation Systems for On-Road Motor Vehicles” (Standard No. J3016-201806 published June 15, 2018, Standard No. J3016-201609 published September 30, 2016, and prior and future versions of this standard). The vehicle 1100 can be capable of implementing functionality that complies with one or more of Levels 3-5 of autonomous driving. For example, depending on the embodiment, the vehicle 1100 can be capable of implementing conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0097] The vehicle 1100 can include components such as a chassis, a body, wheels (e.g., 2, 4, 6, 8, 18, etc.), tires, axles, and other components of a vehicle. The vehicle 1100 can include a propulsion system 1150 such as an internal combustion engine, a hybrid power plant, an all-electric motor, and / or another type of propulsion system. The propulsion system 1150 can be connected to a drivetrain of the vehicle 1100 that can include a transmission in order to effectuate propulsion of the vehicle 1100. The propulsion system 1150 can be controlled in response to receiving a signal from a throttle / accelerator 1152.

[0098] A steering system 1154, which can include a steering wheel, can be used to steer the vehicle 1100 (e.g., along a desired path or route) while the propulsion system 1150 is operating (e.g., while the vehicle is in motion). The steering system 1154 can receive signals from a steering actuator 1156. For full automation (Level 5) functionality, the steering wheel can be optional.

[0099] A braking sensor system 1146 can be used to operate the vehicle brakes in response to receiving signals from a braking actuator 1148 and / or a braking sensor.

[0100] One or more controllers 1136, which can include one or more system on a chip (SoC) 1104 Figure 11C ) and / or one or more GPUs, can provide signals (e.g., representing commands) to one or more components and / or systems of the vehicle 1100. For example, the one or more controllers can send signals to operate the vehicle brakes via one or more braking actuators 1148, to operate the steering system 1154 via one or more steering actuators 1156, to operate the propulsion system 1150 via one or more throttle / accelerator 1152. The one or more controllers 1136 can include one or more on-board (e.g., integrated) computing devices (e.g., supercomputers) that process sensor signals and output operational commands (e.g., signals representing commands) to enable autonomous driving and / or to assist a human driver in driving the vehicle 1100. The one or more controllers 1136 can include a first controller 1136 for autonomous driving functionality, a second controller 1136 for functional safety functionality, a third controller 1136 for artificial intelligence functionality (e.g., computer vision), a fourth controller 1136 for infotainment functionality, a fifth controller 1136 for redundancy in emergency situations, and / or other controllers. In some examples, a single controller 1136 can handle two or more of the above functionalities, two or more controllers 1136 can handle a single functionality, and / or any combination thereof.

[0101] One or more controllers 1136 can provide signals for controlling one or more components and / or systems of the vehicle 1100 in response to sensor data (e.g., sensor inputs) received from one or more sensors. Sensor data can be received from, for example and without limitation, global navigation satellite system sensors 1158 (e.g., global positioning system sensors), RADAR sensors 1160, ultrasonic sensors 1162, LIDAR sensors 1164, inertial measurement unit (IMU) sensors 1166 (e.g., accelerometers, gyroscopes, magnetic compasses, magnetometers, etc.), microphones 1196, stereo cameras 1168, wide-angle cameras 1170 (e.g., fisheye cameras), infrared cameras 1172, surround cameras 1174 (e.g., 360 degree cameras), long and / or medium range cameras 1198, speed sensors 1144 (e.g., to measure the speed of the vehicle 1100), vibration sensors 1142, steering sensors 1140, brake sensors (e.g., as part of a brake sensor system 1146), and / or other sensor types.

[0102] One or more of the controllers 1136 can receive inputs (e.g., represented by input data) from the instrument cluster 1132 of the vehicle 1100 and provide outputs (e.g., represented by output data, display data, etc.) via a human-machine interface (HMI) display 1134, audible annunciators, speakers, and / or via other components of the vehicle 1100. These outputs can include information such as vehicle speed, velocity, time, map data (e.g., HD map 1122), location data (e.g., the location of the vehicle 1100, e.g., on a map), direction, locations of other vehicles (e.g., an occupancy grid), information about objects and object states as perceived by the controllers 1136, and so on. For example, the HMI display 1134 can 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., change lanes now, exit 34B in two miles, etc.). Figure 11C

[0103] The vehicle 1100 also includes a network interface 1124 that can communicate over one or more networks using one or more wireless antennas 1126 and / or modems. For example, the network interface 1124 can be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, etc. The one or more wireless antennas 1126 can also enable communication between objects (e.g., vehicles, mobile devices, etc.) in an implementation environment 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. ​

[0104] Figure 11B FIG. 1 illustrates an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. Figure 11A FIG. 2 illustrates example camera positions and fields of view of the example autonomous vehicle 1100 of FIG. 1. The cameras and respective fields of view are one example embodiment and are not intended to be limiting. For example, additional and / or alternative cameras can be included, and / or the cameras can be located at different positions on the vehicle 1100.

[0105] Camera types for the cameras can include, but are not limited to, digital cameras that can be suitable for use with components and / or systems of the vehicle 1100. The cameras can operate at Automotive Safety Integrity Level (ASIL) B and / or at another ASIL. The camera types can have any image capture rate, such as 60 frames per second (fps), 120 fps, 240 fps, etc., depending on the embodiment. The cameras can be capable of using a rolling shutter, a global shutter, another type of shutter, or a combination thereof. In some examples, a color filter array can 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, clear pixel cameras, such as cameras with a

[0106] 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-functional mono camera can be installed to provide functions including lane departure warning, traffic sign assist, and intelligent headlamp control. One or more (e.g., all) of the cameras can simultaneously record and provide image data (e.g., video).

[0107] One or more of the cameras can be installed in mounting assemblies, such as custom designed (3-D printed) assemblies, in order to cut off stray light and reflections from within the car (e.g., reflections from the dashboard reflected in the windshield mirror) that can interfere with the image data capture capabilities of the cameras. With regard to wing mirror mounting assemblies, the wing mirror assemblies can be custom 3-D printed such that the camera mounting plates match the shape of the wing mirrors. In some examples, one or more cameras can be integrated into the wing mirrors. For side view cameras, one or more cameras can also be integrated into the four pillars at each corner of the cab.

[0108] A camera with a field of view that includes the environment in front of the vehicle 1100 (e.g., a front-facing camera) can be used for surround view to help identify forward paths and obstacles, and, with the assistance of one or more controllers 1136 and / or control SoCs, to provide information crucial for generating an occupancy grid and / or determining a preferred vehicle path. 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 in ADAS functions and systems, including Lane Departure Warning (“LDW”), Autonomous Cruise Control (“ACC”), and / or other functions such as traffic sign recognition.

[0109] A variety of cameras can be used in front-facing configurations, including monocular camera platforms such as CMOS (Complementary Metal-Oxide-Semiconductor) color imagers. Another example could be a wide-angle camera 1170, which can be used to perceive objects entering the field of view from the periphery (e.g., pedestrians, traffic at intersections, or bicycles). Although Figure 11B The image shows only one wide-angle camera, but any number of wide-angle cameras 1170 can be present on vehicle 1100. Furthermore, a remote camera 1198 (e.g., a pair of long-view stereo cameras) can be used for depth-based object detection, especially for objects for which a neural network has not yet been trained. The remote camera 1198 can also be used for object detection and classification, as well as basic object tracking.

[0110] One or more stereo cameras 1168 may also be included in a front-mounted configuration. The stereo camera 1168 may include an integrated control unit comprising a scalable processing unit that can provide a multi-core microprocessor and programmable logic (FPGA) with an integrated CAN or Ethernet interface on a single chip. Such a unit can be used to generate a 3D map of the vehicle environment, including distance estimates for all points in the image. Alternative stereo cameras 1168 may include a compact stereo vision sensor that may include two camera lenses (one on each side) and an image processing chip capable of measuring the distance from the vehicle to a target object and using the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 1168 may be used in addition to those described herein, or alternatively.

[0111] A camera (e.g., a side-view camera) having a field of view that includes the side of the vehicle 1100 can be used for surround view, providing information for creating and updating occupancy grids and generating side-impact collision warnings. For example, a surround camera 1174 (e.g., ...) Figure 11BFour surround cameras 1174) can be placed on the vehicle 1100. The surround cameras 1174 can include wide-view cameras 1170, fisheye cameras, 360-degree cameras, and / or the like. In one example, four fisheye cameras can be placed on the front, back, and sides of the vehicle. In an alternative arrangement, the vehicle can use three surround cameras 1174 (e.g., left, right, and back) and can utilize one or more other cameras (e.g., a forward-facing camera) as a fourth surround view camera.

[0112] Cameras with a field of view that includes an environmental portion of the back of the vehicle 1100 (e.g., rearview cameras) can be used to assist with parking, surround view, back collision warnings, and creating and updating an occupancy grid. A wide variety of cameras can be used, including but not limited to cameras that are also suitable as front-facing cameras (e.g., long and / or mid-range cameras 1198, stereo cameras 1168, infrared cameras 1172, etc.) as described herein.

[0113] Figure 11C FIG. 1 illustrates an example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure. Figure 11A FIG. 1 illustrates an example autonomous vehicle 1100 for use in accordance with some embodiments of the present disclosure.

[0114] Figure 11C Each of the components, features, and systems of the vehicle 1100 in FIG. 1 are illustrated as being connected via a bus 1102. The bus 1102 can include a controller area network (CAN) data interface (alternatively referred to herein as a "CAN bus"). The CAN can be a network within the vehicle 1100 that is used to assist in controlling various features and functions of the vehicle 1100, such as the actuation of brakes, acceleration, braking, steering, windshield wipers, etc. The CAN bus can be configured to have tens or even hundreds of nodes, each with its own unique identifier (e.g., CAN ID). The CAN bus can be read to find steering wheel angle, ground speed, engine revolutions per minute (RPM), button positions, and / or other vehicle status indicators. The CAN bus can be ASIL B compliant.

[0115] Although bus 1102 is described herein as a CAN bus, this is not intended to be limiting. For example, FlexRay and / or Ethernet can be used in addition to or instead of a CAN bus. Further, although bus 1102 is represented with a single line, this is not intended to be limiting. For example, there can be any number of buses 1102, which can include one or more CAN buses, one or more FlexRay buses, one or more Ethernet buses, and / or one or more other types of buses that use different protocols. In some examples, two or more buses 1102 can be used to perform different functions, and / or can be used for redundancy. For example, a first bus 1102 can be used for collision avoidance functions, and a second bus 1102 can be used for drive control. In any example, each bus 1102 can communicate with any component of vehicle 1100, and two or more buses 1102 can communicate with the same components. In some examples, each SoC 1104, each controller 1136, and / or each computer within the vehicle can have access to the same input data (e.g., input from sensors of vehicle 1100), and can be connected to a common bus, such as a CAN bus.

[0116] Vehicle 1100 can include one or more controllers 1136, such as those described herein with respect to Figure 11A controllers. Controllers 1136 can be used for a wide variety of functions. Controllers 1136 can be coupled to any other different components and systems of vehicle 1100, and can be used for control of vehicle 1100, artificial intelligence of vehicle 1100, infotainment for vehicle 1100, and / or the like.

[0117] Vehicle 1100 can include one or more system on chips (SoCs) 1104. SoCs 1104 can include CPUs 1106, GPUs 1108, processors 1110, caches 1112, accelerators 1114, data stores 1116, and / or other components and features not illustrated. In a wide variety of platforms and systems, SoCs 1104 can be used to control vehicle 1100. For example, one or more SoCs 1104 can be used in a system (e.g., a system of vehicle 1100) in conjunction with HD map 1122, which can obtain map refreshes and / or updates from one or more servers (e.g., one or more servers 1178) via network interface 1124. Figure 11D

[0118] ​CPU 1106 can include a CPU cluster or CPU complex (alternatively referred to herein as a“CCPLEX”). CPU 1106 can include multiple cores and / or L2 caches. For example, in some embodiments, CPU 1106 can include eight cores in a coherent multi-processor configuration. In some embodiments, CPU 1106 can include four dual-core clusters, with each cluster having a dedicated L2 cache (e.g., a 2 MB L2 cache). CPU 1106 (e.g., the CCPLEX) can be configured to support simultaneous cluster operation, such that any combination of clusters of CPU 1106 can be active at any given time.

[0119] CPU 1106 can implement power management capabilities including one or more of the following features: individual hardware blocks can be automatically clock-gated when idle to save dynamic power; each core clock can be gated when the core is not actively executing instructions due to execution of WFI / WFE instructions; each core can be independently power-gated; each core cluster can be independently clock-gated when all cores are clock-gated or power-gated; and / or each core cluster can be independently power-gated when all cores are power-gated. CPU 1106 can further implement enhanced algorithms for managing power states, with specified allowed power states and desired wake-up times, and the hardware / microcode determines the optimal power state for the cores, clusters, and CCPLEX to enter. The processing cores can support a simplified power state entry sequence in software, with the work offloaded to microcode.

[0120] GPU 1108 can include an integrated GPU (alternatively referred to herein as an“iGPU”). GPU 1108 can be programmable and efficient for parallel workloads. In some examples, GPU 1108 can use an enhanced tensor instruction set. GPU 1108 can include one or more streaming microprocessors, where each streaming microprocessor can include an Ll cache (e.g., an Ll cache having at least 96 KB of storage capacity), and two or more of the streaming microprocessors can share an L2 cache (e.g., an L2 cache having 512 KB of storage capacity). In some embodiments, GPU 1108 can include at least eight streaming microprocessors. GPU 1108 can use a compute application programming interface (API). In addition, GPU 1108 can use one or more parallel computing platforms and / or programming models (e.g., NVIDIA’s CUDA).

[0121] In the case of automotive and embedded uses, the GPU 1108 can be power-optimized for best performance. For example, the GPU 1108 can be fabricated on a fin field-effect transistor (FinFET) for lower power consumption. However, this is not intended to be limiting, and the GPU 1108 can be fabricated using other semiconductor manufacturing processes. Each streaming microprocessor can incorporate 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, an L0 instruction cache, a thread warp scheduler, a dispatch unit, and / or a 64 KB register file. Further, the streaming microprocessor can include independent parallel integer and floating point data paths to exploit the mix of computation and address computation for efficient execution of workloads. The streaming microprocessor can include independent thread scheduling capabilities to allow for finer-grain synchronization and cooperation between parallel threads. The streaming microprocessor can include a combined LI data cache and shared memory unit to improve performance while simplifying programming.

[0122] The GPU 1108 can include a high bandwidth memory (HBM) and / or a 16 GB HBM2 memory subsystem that provides approximately 900 GB / s of peak memory bandwidth in some examples. In some examples, in addition to or alternatively from HBM memory, synchronous graphics random access memory (SGRAM) can be used, such as fifth generation graphics double data rate synchronous random access memory (GDDR5).

[0123] The GPU 1108 can include a unified memory technology that includes an access counter to allow memory pages to be migrated more precisely to the processors that access them most frequently, improving efficiency of memory ranges shared between processors. In some examples, address translation services (ATS) support can be used to allow the GPU 1108 to access CPU 1106 page tables directly. In such examples, when the GPU 1108 memory management unit (MMU) experiences a miss, an address translation request can be transmitted to the CPU 1106. In response, the CPU 1106 can look up the virtual-to-physical mapping for the address in its page tables and transmit the translation back to the GPU 1108. In this way, the unified memory technology can allow a single unified virtual address space for memory of both the CPU 1106 and the GPU 1108, simplifying GPU 1108 programming and porting applications to the GPU 1108.

[0124] In addition, GPU 1108 can include an access counter that can track how often GPU 1108 accesses memory of other processors. The access counter can help ensure that memory pages are moved to the physical memory of the processor that most frequently accesses those pages.

[0125] SoC 1104 can include any number of caches 1112, including those described herein. For example, caches 1112 can include an L3 cache that is available to both CPU 1106 and GPU 1108 (e.g., connected to both CPU 1106 and GPU 1108). Caches 1112 can include a write-back cache that can track the state of a line, for example, by using a cache coherency protocol (e.g., MEI, MESI, MSI, etc.). Depending on the embodiment, the L3 cache can include 4MB or more, although smaller cache sizes can also be used.

[0126] SoC 1104 can include an arithmetic logic unit (ALU) that can be utilized in processing any of a variety of tasks or operations with respect to vehicle 1100, such as processing a DNN. In addition, SoC 1104 can include a floating point unit (FPU) (or other mathematical co-processor or digital co-processor type) for performing mathematical operations within the system. For example, SoC 104 can include one or more FPUs integrated as execution units within CPU 1106 and / or GPU 1108.

[0127] SoC 1104 can include one or more accelerators 1114 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, SoC 1104 can include a hardware accelerator cluster that can include optimized hardware accelerators and / or a large on-chip memory. The large on-chip memory (e.g., 4MB SRAM) can enable the hardware accelerator cluster to accelerate neural networks and other computations. The hardware accelerator cluster can be used to supplement GPU 1108 and offload some of the tasks of GPU 1108 (e.g., freeing up more cycles of GPU 1108 for performing other tasks). As one example, accelerators 1114 can be used for targeted workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are stable enough to accelerate easily. As used herein, the term “CNN” can include all types of CNNs, including region-based or region convolutional neural networks (RCNNs) and fast RCNNs (e.g., for object detection).

[0128] The accelerators 1114 (e.g., a cluster of hardware accelerators) can include a deep learning accelerator (DLA). The DLA can include one or more tensor processing units (TPUs) that can be configured to provide an additional 100 billion operations per second for deep learning applications and inferencing. The TPU can be an accelerator that is configured to perform image processing functions (e.g., for CNNs, RCNNs, etc.) and is optimized for performing image processing functions. The DLA can be further optimized for a specific set of neural network types and floating point operations and inferencing. The design of the DLA can provide higher performance per mm than general purpose GPUs and far exceeds the performance of CPUs. The TPU can perform several functions including single instance convolution functions, support for INT8, INT16, and FP16 data types for both features and weights, for example, and post-processor functions.

[0129] The DLA can perform neural networks, especially CNNs, on processed or unprocessed data for any of a wide variety of functions, such as and not by way of limitation: CNNs for object recognition and detection using data from camera sensors; CNNs for distance estimation using data from camera sensors; CNNs for emergency vehicle detection and identification and detection using data from microphones; CNNs for facial recognition and vehicle owner identification using data from camera sensors; and / or CNNs for safety and / or safety related events.

[0130] The DLA can perform any of the functions of the GPU 1108, and by using an inferencing accelerator, the designer can target the DLA or the GPU 1108 for any function, for example. For example, the designer can focus the processing and floating point operations of CNNs on the DLA and leave other functions to the GPU 1108 and / or other accelerators 1114.

[0131] The accelerators 1114 (e.g., a cluster of hardware accelerators) can include a programmable vision accelerator (PVA), which can be alternatively referred to herein as a computer vision accelerator. The PVA can 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 can provide a balance between performance and flexibility. For example, each PVA can include any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors, for example and not by way of limitation.

[0132] The RISC cores can interact with image sensors (e.g., image sensors of any of the cameras described herein), image signal processors, and / or the like. Each of these RISC cores can include any number of memories. Depending on the embodiment, the RISC cores can use any of several protocols. In some examples, the RISC cores can execute a real-time operating system (RTOS). The RISC cores can be implemented using one or more integrated circuit devices, application specific integrated circuits (ASICs), and / or memory devices. For example, the RISC cores can include instruction caches and / or tightly coupled RAM.

[0133] The DMA can enable components of the PVA to access system memory independently of the CPU 1106. The DMA can support any number of features to provide optimization to the PVA, including but not limited to supporting multi-dimensional addressing and / or circular addressing. In some examples, the DMA can support addressing up to six or more dimensions, which can include block width, block height, block depth, horizontal block step, vertical block step, and / or depth step.

[0134] The vector processor can be a programmable processor that can be designed to efficiently and flexibly execute programming for computer vision algorithms and provide signal processing capabilities. In some examples, the PVA can include a PVA core and two vector processing subsystem partitions. The PVA core can include a processor subsystem, one or more DMA engines (e.g., two DMA engines), and / or other peripherals. The vector processing subsystems can operate as the main processing engines 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), very long instruction word (VLIW) digital signal processor. The combination of SIMD and VLIW can enhance throughput and speed.

[0135] Each of the vector processors can include an instruction cache and can be coupled to a dedicated memory. As a result, in some examples, each of the vector processors can be configured to execute independently of the other vector processors. In other examples, the vector processors included in a particular PVA can be configured to employ data parallelization. For example, in some embodiments, multiple vector processors included in a single PVA can execute the same computer vision algorithm, but on different regions of an image. In other examples, the vector processors included in a particular PVA can execute different computer vision algorithms on the same image simultaneously, or even different algorithms on a sequence of images or portions of an image. Any number of PVAs can be included in a hardware accelerator cluster, and any number of vector processors can be included in each of those PVAs, among other things. Furthermore, the PVAs can include additional error-correcting code (ECC) memory to enhance overall system security.

[0136] The accelerator 1114 (e.g., hardware accelerator cluster) can include on-chip computer vision networks and SRAM to provide high bandwidth, low latency SRAM for the accelerator 1114. In some examples, the on-chip memory can include at least 4 MB of SRAM composed of, for example and without limitation, eight field-programmable memory blocks, which can be accessed by both the PVA and the DLA. Each pair of memory blocks can include an advanced peripheral bus (APB) interface, configuration circuitry, a controller, and a multiplexer. Any type of memory can be used. The PVA and the DLA can access the memory via a backbone that provides high-speed memory access to the PVA and the DLA. The backbone can include on-chip computer vision networks that interconnect the PVA and the DLA to the memory (e.g., using APB).

[0137] The on-chip computer vision networks can include an interface that determines that both the PVA and the DLA provide ready and valid signals before transmitting any control signals / addresses / data. Such an interface can provide separate phases and separate channels for transmitting control signals / addresses / data, as well as burst communications for continuous data transmission. This type of interface can comply with ISO 26262 or IEC 61508 standards, but other standards and protocols can also be used.

[0138] In some examples, the SoC 1104 can include a real-time ray tracing hardware accelerator, such as described in U.S. Patent Application No. 16 / 101,232, filed August 10, 2018. The real-time ray tracing hardware accelerator can be used to quickly and efficiently determine locations and extents of objects (e.g., within a world model) in order to generate real-time visualizations simulations for RADAR signal interpretation, for sound propagation synthesis and / or analysis, for SONAR system simulation, for general wave propagation simulation, for comparison with LIDAR data for purposes of localization and / or other functions, and / or for other uses. In some embodiments, one or more tree traversal units (TTUs) can be used to perform one or more ray tracing related operations.

[0139] The accelerator 1114 (e.g., a hardware accelerator cluster) has a wide range of uses for autonomous driving. The PVA can be a programmable vision accelerator that can be used for key processing stages in ADAS and autonomous vehicles. The capabilities of the PVA are a good match for algorithm domains that require predictable processing, low power, and low latency. In other words, the PVA performs well on semi-dense or dense regular computations, and even on small data sets that require predictable runtimes with low latency and low power. Thus, in the context of a platform for autonomous vehicles, the PVA is designed to run classical computer vision algorithms because they are effective at object detection and integer math operations.

[0140] For example, according to one embodiment of the technology, the PVA is used to perform computer stereo vision. In some examples, a semi-global matching based algorithm can be used, although 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.). The PVA can perform computer stereo vision functions on input from two monocular cameras.

[0141] In some examples, the PVA can be used to perform dense optical flow. Raw RADAR data is processed according to a process (e.g., using a 4D fast Fourier transform) to provide processed RADAR. In other examples, the PVA is used for time-of-flight depth processing, such as by processing raw time-of-flight data to provide processed time-of-flight data.

[0142] The DLA can be used to run any type of network to enhance control and driving safety, including, for example, a neural network that outputs a confidence metric for each object detection. Such a confidence value can be interpreted as a probability, or as providing a relative "weight" for each detection compared to other detections. The confidence value enables the system to make further decisions about which detections should be considered true positive detections and not 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, false positive detections would cause the vehicle to automatically perform an emergency brake, which is obviously undesirable. Thus, only the most confident detections should be considered a trigger for AEB. The DLA can run a neural network for regression of a confidence value. The neural network can take as its input at least some subset of parameters, such as a bounding box dimension, a ground plane estimate obtained (e.g., from another subsystem), inertial measurement unit (IMU) sensor 1166 outputs related to vehicle 1100 orientation, distance, 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., LIDAR sensor 1164 or RADAR sensor 1160), etc.

[0143] SoC 1104 can include one or more data stores 1116 (e.g., memory). Data stores 1116 can be on-chip memory of SoC 1104, which can store neural networks to be executed on the GPU and / or DLA. In some examples, for redundancy and safety, data stores 1116 can be large enough in capacity to store multiple instances of a neural network. Data stores 1112 can include L2 or L3 cache 1112. References to data stores 1116 can include references to memory associated with PVA, DLA, and / or other accelerators 1114 as described herein.

[0144] SoC 1104 can include one or more processors 1110 (e.g., embedded processors). The processors 1110 can include a boot and power management processor, which can be a specialized processor and subsystem for handling boot power and management functions, as well as security implementation. The boot and power management processor can be part of the SoC 1104 boot sequence and can provide runtime power management services. The boot power and management processor can provide clock and voltage programming, auxiliary system low power state transitions, SoC 1104 thermal and temperature sensor management, and / or SoC 1104 power state management. Each temperature sensor can be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 1104 can use the ring oscillator to detect the temperature of the CPU 1106, GPU 1108, and / or accelerator 1114. If it is determined that the temperature exceeds a threshold, the boot and power management processor can enter a temperature fault routine and place the SoC 1104 in a lower power state and / or place the vehicle 1100 in a driver safe park mode (e.g., safely park the vehicle 1100).

[0145] The processors 1110 can also include a set of embedded processors that can be used as an audio processing engine. The audio processing engine can be an audio subsystem that allows for full hardware support for multi-channel audio over multiple interfaces, as well as a range of extensive and flexible audio I / O interfaces. In some examples, the audio processing engine is a specialized processor core with a digital signal processor with dedicated RAM.

[0146] The processors 1110 can also include an always-on processor engine, which can provide the necessary hardware features to support low-power sensor management and wake-up use cases. The always-on processor engine can include a processor core, tightly coupled RAM, supporting peripherals (e.g., timers and interrupt controllers), various I / O controller peripherals, and routing logic.

[0147] The processors 1110 can also include a security cluster engine, which includes a specialized processor subsystem that handles security management for automotive applications. The security cluster engine can include two or more processor cores, tightly coupled RAM, supporting peripherals (e.g., timers, interrupt controllers, etc.), and / or routing logic. In a secure mode, the two or more cores can operate in a lockstep mode and act as a single core with comparison logic that detects any differences between their operations.

[0148] The processors 1110 can also include a real-time camera engine, which can include a specialized processor subsystem for handling real-time camera management.

[0149] The processor 1110 can further include a high dynamic range signal processor, which can include an image signal processor, which is a hardware engine that is part of the camera processing pipeline.

[0150] The processor 1110 can include a video image compositor, which can be a processing block (e.g., implemented on a microprocessor), that implements video post-processing functions needed by the video playback application to produce the final image for the player window. The video image compositor can perform lens distortion correction on the wide-angle camera 1170, the surround camera 1174, and / or on the cab-in monitor camera sensors. The cab-in monitor camera sensors are preferably monitored by a neural network running on another instance of the advanced SoC, configured to recognize cab-in events and respond accordingly. The cab-in system can perform lip reading to activate mobile phone services and place a call, dictate an email, change the vehicle destination, 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 otherwise.

[0151] The video image compositor can include enhanced temporal noise reduction for spatial and temporal noise reduction. For example, where motion is present in the video, the noise reduction appropriately weights the spatial information, reducing the weight of information provided by neighboring frames. Where the image or portions of the image do not include motion, the temporal noise reduction performed by the video image compositor can use information from previous images to reduce noise in the current image.

[0152] The video image compositor can also be configured to perform stereo correction on input stereo lens frames. The video image compositor can further be used for user interface composition when the operating system desktop is in use and the GPU 1108 does not need to continuously render new surfaces. Even when the GPU 1108 is powered on and active, doing 3D rendering, the video image compositor can be used to offload the GPU 1108 to improve performance and responsiveness.

[0153] The SoC 1104 can further include a Mobile Industry Processor Interface (MIPI) camera serial interface for receiving video and input from the cameras, a high-speed interface, and / or a video input block that can be used for camera and related pixel input functions. The SoC 1104 can further include an input / output controller that can be controlled by software and can be used to receive I / O signals that are not committed to a particular role.

[0154] The SoC 1104 can also include a wide range of peripheral device interfaces to enable communication with peripherals, audio codecs, power management, and / or other devices. The SoC 1104 can be used to process data from cameras (connected over Gigabit Multimedia Serial Link and Ethernet), sensors (e.g., LIDAR sensor 1164, RADAR sensor 1160, etc. that can be connected over Ethernet), data from the bus 1102 (e.g., speed of the vehicle 1100, steering wheel position, etc.), data from GNSS sensor 1158 (connected over Ethernet or CAN bus). The SoC 1104 can also include dedicated high performance mass storage controllers that can include their own DMA engines and that can be used to free up the CPU 1106 from routine data management tasks.

[0155] The SoC 1104 can be an end-to-end platform with a flexible architecture that spans automation levels 3-5, providing an integrated functional safety architecture for a platform that leverages and efficiently uses computer vision and ADAS technology to achieve diversity and redundancy, along with deep learning tools. The SoC 1104 can be faster, more reliable, and even more energy and space efficient than conventional systems. For example, the accelerators 1114, when combined with the CPU 1106, GPU 1108, and data storage 1116, can provide a fast and efficient platform for level 3-5 autonomous vehicles.

[0156] The technology thus provides capabilities and functionality that cannot be achieved with conventional systems. For example, computer vision algorithms can be executed on CPUs that can be configured using high-level programming languages such as the C programming language to perform a wide variety of processing algorithms across a wide variety of visual data. However, CPUs often cannot meet the performance requirements of many computer vision applications, such as those related to, for example, execution time and power consumption. In particular, many CPUs cannot execute complex object detection algorithms in real time, which is a requirement for on-board ADAS applications and for practical level 3-5 autonomous vehicles.

[0157] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a cluster of hardware accelerators, the technology described herein allows multiple neural networks to be executed simultaneously and / or sequentially, and the results to be combined together to achieve level 3-5 autonomous driving functionality. For example, a CNN executed on a DLA or dGPU (e.g., GPU 1120) can include text and word recognition, allowing a supercomputer to read and understand traffic signs, including signs for which a neural network has not been specifically trained. The DLA can also include a neural network that is able to recognize, interpret, and provide a semantic understanding of the sign, and pass that semantic understanding to a path planning module running on the CPU complex.

[0158] As another example, multiple neural networks can be run simultaneously as required for level 3, 4, or 5 driving. For example, a warning sign consisting of the words "Caution: flashing lights indicate icy conditions" along with electric lights can be interpreted by several neural networks independently or collectively. The sign itself can be recognized by a first deployed neural network (e.g., a trained neural network) as a traffic sign, the text "flashing lights indicate icy conditions" can be interpreted by a second deployed neural network that informs the vehicle's path planning software (preferably executing on the CPU complex) that icy conditions exist when flashing lights are detected. The flashing lights can be recognized by operating a third deployed neural network over multiple frames that informs the vehicle's path planning software of the presence (or absence) of flashing lights. All three neural networks can be run simultaneously, for example, within the DLA and / or on the GPU 1108.

[0159] In some examples, a CNN for face recognition and owner recognition can use data from the camera sensors to recognize the presence of an authorized driver and / or owner of the vehicle 1100. A processing engine always on the sensor can be used to unlock the vehicle and turn on the lights when the owner approaches the driver's door, and in a safe mode, disable the vehicle when the owner leaves the vehicle. In this way, the SoC 1104 provides security against theft and / or carjacking.

[0160] In another example, a CNN for emergency vehicle detection and recognition can use data from the microphones 1196 to detect and recognize emergency vehicle sirens. In contrast to conventional systems that use a general classifier to detect sirens and manually extract features, the SoC 1104 uses a CNN to classify ambient and urban sounds as well as to classify visual data. In a preferred embodiment, a CNN running on the DLA is trained to recognize the relative closing speed of an emergency vehicle (e.g., by using the Doppler effect). The CNN can also be trained to recognize emergency vehicles specific to the local area in which the vehicle is operating as recognized by the GNSS sensor 1158. 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 recognize sirens that are only North American. Once an emergency vehicle is detected, a control program can be used to execute an emergency vehicle safety routine, slow the vehicle down, pull over to the side of the road, stop the vehicle, and / or idle the vehicle until the emergency vehicle passes, with the assistance of the ultrasonic sensors 1162.

[0161] The vehicle can include a CPU 1118 (e.g., a discrete CPU or dCPU) that can be coupled to the SoC 1104 via a high-speed interconnect (e.g., PCIe). The CPU 1118 can include, for example, an X86 processor. The CPU 1118 can be used to perform any of a wide variety of functions, including, for example, arbitrating potentially inconsistent results between ADAS sensors and the SoC 1104, and / or monitoring the status and health of the controller 1136 and / or infotainment SoC 1130.

[0162] The vehicle 1100 can include a GPU 1120 (e.g., a discrete GPU or dGPU) that can be coupled to the SoC 1104 via a high-speed interconnect (e.g., NVIDIA’s NVLINK). The GPU 1120 can provide additional artificial intelligence functionality, for example, by executing redundant and / or different neural networks, and can be used to train and / or update neural networks based at least in part on input (e.g., sensor data) from sensors of the vehicle 1100.

[0163] The vehicle 1100 can also include a network interface 1124 that can include one or more wireless antennas 1126 (e.g., one or more wireless antennas for different communication protocols, such as cellular antennas, Bluetooth antennas, etc.). The network interface 1124 can be used to enable wireless connections over the Internet with a cloud (e.g., with the server 1178 and / or other network devices), with other vehicles, and / or with computing devices (e.g., client devices of passengers). For communication 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 car-to-car communication link. The car-to-car communication link can provide the vehicle 1100 with information about vehicles that are approaching the vehicle 1100 (e.g., vehicles in front of, to the side of, and / or behind the vehicle 1100). This functionality can be part of a cooperative adaptive cruise control functionality of the vehicle 1100.

[0164] The network interface 1124 can include a SoC that provides modulation and demodulation functionality and enables the controller 1136 to communicate over a wireless network. The network interface 1124 can 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 can be performed through well-known processes and / or can be performed using a super-heterodyne process. In some examples, the radio frequency front end functionality can be provided by a separate chip. The network interface can include wireless functionality for communication over LTE, WCDMA, UMTS, GSM, CDMA2000, Bluetooth, Bluetooth LE, Wi-Fi, Z-Wave, ZigBee, LoRaWAN, and / or other wireless protocols.

[0165] The vehicle 1100 can also include a data store 1128, which can include off-chip (e.g., off-SoC 1104) storage. The data store 1128 can include one or more storage elements, including RAM, SRAM, DRAM, VRAM, flash memory, hard disks, and / or other components and / or devices that can store data for at least one bit.

[0166] The vehicle 1100 can also include a GNSS sensor 1158. The GNSS sensor 1158 (e.g., GPS, assisted GPS sensor, differential GPS (DGPS) sensor, etc.) is used to assist in mapping, perception, occupancy grid generation, and / or path planning functions. Any number of GNSS sensors 1158 can be used, including, for example and without limitation, a GPS using a USB connector with an Ethernet-to-serial (RS-232) bridge.

[0167] The vehicle 1100 can also include a RADAR sensor 1160. The RADAR sensor 1160 can be used by the vehicle 1100 for long-range vehicle detection, even in darkness and / or adverse weather conditions. The RADAR functional safety level can be ASIL B. The RADAR sensor 1160 can use the CAN and / or the bus 1102 (e.g., to transmit data generated by the RADAR sensor 1160) for control as well as access to object tracking data, in some examples, Ethernet for access to raw data. A wide variety of RADAR sensor types can be used. For example and without limitation, the RADAR sensor 1160 can be suitable for front, rear, and side RADAR use. In some examples, a pulsed Doppler RADAR sensor is used.

[0168] The RADAR sensor 1160 can include different configurations, such as long-range with narrow field of view, short-range with wide field of view, short-range side coverage, and so on. In some examples, long-range RADAR can be used for adaptive cruise control functionality. Long-range RADAR systems can provide a wide field of view (e.g., 250 m range) achieved through two or more independent scans. The RADAR sensor 1160 can help distinguish between static and moving objects, and can be used by the ADAS system for emergency brake assist and forward collision warning. The long-range RADAR sensor can include a single-station multi-mode RADAR with multiple (e.g., six or more) fixed RADAR antennas, as well as a high-speed CAN and FlexRay interface. In examples with six antennas, the central four antennas can create focused beam patterns designed to record the surroundings of the vehicle 1100 at higher speed with minimal traffic interference from adjacent lanes. The other two antennas can extend the field of view, making it possible to quickly detect vehicles entering or leaving the lane of the vehicle 1100.

[0169] As one example, a mid-range RADAR system can include a range of up to 1160 m (front) or 80 m (rear) and a field of view of up to 42 degrees (front) or 1150 degrees (rear). A short-range RADAR system can include, but is not limited to, RADAR sensors designed to be mounted at both ends of the rear bumper. When mounted at both ends of the rear bumper, such a RADAR sensor system can create two beams that continuously monitor the rear and the blind spot next to the vehicle.

[0170] A short-range RADAR system can be used in an ADAS system for blind spot detection and / or lane change assist.

[0171] The vehicle 1100 can also include ultrasonic sensors 1162. Ultrasonic sensors 1162 that can be placed on the front, rear, and / or sides of the vehicle 1100 can be used for parking assist and / or to create and update an occupancy grid. A wide variety of ultrasonic sensors 1162 can be used, and different ultrasonic sensors 1162 can be used for different detection ranges (e.g., 2.5 m, 4 m). The ultrasonic sensors 1162 can operate at an ASIL B functional safety level.

[0172] The vehicle 1100 can include LIDAR sensors 1164. The LIDAR sensors 1164 can be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensors 1164 can be at an ASIL B functional safety level. In some examples, the vehicle 1100 can include multiple LIDAR sensors 1164 (e.g., two, four, six, etc.) that can use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0173] In some examples, the LIDAR sensors 1164 can be capable of providing a list of objects and their distances for a 360-degree field of view. A commercially available LIDAR sensor 1164 can have, for example, an advertised range of approximately 1100 m, a precision of 2 cm - 3 cm, and support for a 1100 Mbps Ethernet connection. In some examples, one or more flush-mounted LIDAR sensors 1164 can be used. In such examples, the LIDAR sensors 1164 can be implemented as small devices that can be embedded into the front, rear, sides, and / or corners of the vehicle 1100. In such examples, the LIDAR sensors 1164 can provide a field of view of up to 120 degrees horizontal and 35 degrees vertical for low reflectivity objects, with a range of 200 m. Front-mounted LIDAR sensors 1164 can be configured for a horizontal field of view between 45 degrees and 135 degrees.

[0174] In some examples, LIDAR technology such as 3D Flash LIDAR can also be used. 3D Flash LIDAR uses a flash of laser light as a source of emission to illuminate the vehicle’s surroundings up to about 200 m. The flash LIDAR unit includes a receptor 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 for the generation of highly accurate and distortion-free images of the surroundings with each laser flash. In some examples, four flash LIDAR sensors can be deployed, one on each side of the vehicle 1100. Available 3D flash LIDAR systems include solid-state 3D staring array LIDAR cameras (e.g., non-scanning LIDAR devices) that have no moving parts other than a fan. The flash LIDAR device can use 5 nanosecond Class I (eye-safe) laser pulses per frame and can capture the reflected laser light in the form of 3D range point clouds 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 1164 can be less susceptible to motion blur, vibration, and / or jostling.

[0175] The vehicle can also include an IMU sensor 1166. In some examples, the IMU sensor 1166 can be located at the center of the rear axle of the vehicle 1100. The IMU sensor 1166 can include, for example and without limitation, an accelerometer, a magnetometer, a gyroscope, a magnetic compass, and / or other sensor types. In some examples, such as in six-axis applications, the IMU sensor 1166 can include an accelerometer and a gyroscope, while in nine-axis applications, the IMU sensor 1166 can include an accelerometer, a gyroscope, and a magnetometer.

[0176] In some embodiments, the IMU sensor 1166 can be implemented as a microelectromechanical systems (MEMS) based inertial navigation system (INS) that combines a microelectromechanical systems (MEMS) inertial sensor, a high-sensitivity GPS receiver, and advanced Kalman filtering algorithms to provide estimates of position, velocity, and attitude. As such, in some examples, the IMU sensor 1166 can enable the vehicle 1100 to estimate heading without input from a magnetic sensor by directly observing the change in velocity from GPS to the IMU sensor 1166 and correlating it. In some examples, the IMU sensor 1166 and the GNSS sensor 1158 can be combined into a single integrated unit.

[0177] The vehicle can include a microphone 1196 placed in and / or around the vehicle 1100. The microphone 1196 can be used for emergency vehicle detection and identification, among other things.

[0178] The vehicle can also include any number of camera types, including stereo cameras 1168, wide-view cameras 1170, infrared cameras 1172, surround-view cameras 1174, long and / or mid-range cameras 1198, and / or other camera types. These cameras can be used to capture image data around the entire periphery of the vehicle 1100. The types of cameras used depend on the embodiment and requirements of the vehicle 1100, and any combination of camera types can be used to provide the necessary coverage around the vehicle 1100. Further, the number of cameras can vary depending on the embodiment. For example, the vehicle can include six cameras, seven cameras, ten cameras, twelve cameras, and / or another number of cameras. As one example and not by way of limitation, the cameras can support Gigabit Multimedia Serial Link (GMSL) and / or Gigabit Ethernet. Each of the cameras is described in more detail herein with respect to Figure 11A and Figure 11B are described in more detail.

[0179] The vehicle 1100 can also include vibration sensors 1142. The vibration sensors 1142 can measure vibrations of components of the vehicle, such as axles. For example, changes in vibration can indicate changes in the road surface. In another example, when two or more vibration sensors 1142 are used, differences between the vibrations can be used to determine the friction or slip of the road surface (e.g., when there is a difference in vibration between a power driven axle and a free spinning axle).

[0180] The vehicle 1100 can include an ADAS system 1138. In some examples, the ADAS system 1138 can include a SoC. The ADAS system 1138 can include adaptive / automatic / autonomous cruise control (ACC), cooperative adaptive cruise control (CACC), forward collision warning (FCW), automatic emergency braking (AEB), lane departure warning (LDW), lane keep assist (LKA), blind spot warning (BSW), rear cross-traffic warning (RCTW), collision warning system (CWS), lane centering (LC), and / or other features and functionality.

[0181] The ACC system can use RADAR sensors 1160, LIDAR sensors 1164, and / or cameras. The ACC system can include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle immediately ahead of the vehicle 1100 and automatically adjusts the vehicle speed to maintain a safe distance from the vehicle ahead. Lateral ACC performs distance keeping and, if necessary, suggests a lane change for the vehicle 1100. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0182] CACC uses information from other vehicles, which can be received from other vehicles indirectly via a wireless link or through a network connection (e.g., through the Internet) via the network interface 1124 and / or the wireless antenna 1126. Direct links can be provided by a vehicle-to-vehicle (V2V) communication link, while indirect links can be an infrastructure-to-vehicle (I2V) communication link. Generally, the V2V communication concept provides information about the immediately preceding vehicles (e.g., vehicles immediately ahead of and in the same lane as the vehicle 1100), while the I2V communication concept provides information about traffic further ahead. A CACC system can include either or both of I2V and V2V information sources. Given information about vehicles ahead of the vehicle 1100, CACC can be more reliable, and it has the potential to improve traffic flow and reduce road congestion.

[0183] FCW systems are designed to alert the driver to a hazard so that the driver can take corrective action. FCW systems use a front-facing camera and / or RADAR sensor 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as displays, speakers, and / or vibrating components. FCW systems can provide warnings in the form of, for example, sound, visual warnings, vibrations, and / or quick brake pulses.

[0184] AEB systems detect 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. AEB systems can use a front-facing camera and / or RADAR sensor 1160 coupled to a dedicated processor, DSP, FPGA, and / or ASIC. When an AEB system detects a hazard, it typically first alerts the driver to take corrective action to avoid a 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 a predicted collision. AEB systems can include technologies such as dynamic brake support and / or crash imminent braking.

[0185] LDW systems provide visual, audible, and / or tactile warnings such as steering wheel or seat vibrations to alert the driver when the vehicle 1100 is crossing lane markers. The LDW system is not activated when the driver indicates an intentional lane departure by activating a turn signal. LDW systems can use a front-side facing camera coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as displays, speakers, and / or vibrating components.

[0186] An LKA system is a variation of the LDW system. If the vehicle 1100 begins to leave the lane, the LKA system provides a steering input or brake to correct the vehicle 1100.

[0187] A BSW system detects and warns the driver of vehicles in the car's blind spot. The BSW system can provide visual, audible, and / or tactile alerts to indicate that merging or changing lanes is unsafe. The system can provide additional warnings when the driver uses a turn signal. The BSW system can use rear side-facing cameras and / or RADAR sensors 1160 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.

[0188] A RCTW system can provide visual, audible, and / or tactile notifications when objects are detected outside the range of the rear-facing camera while the vehicle 1100 is backing up. Some RCTW systems include AEB to ensure that vehicle brakes are applied to avoid a collision. The RCTW system can use one or more rear-facing RADAR sensors 1160 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.

[0189] Conventional ADAS systems can be prone to false positive results, which can annoy and distract the driver, but typically are not catastrophic because the ADAS system alerts the driver and allows the driver to decide whether the safety condition is truly present and act accordingly. However, in an autonomous vehicle 1100, in the case of conflicting results, the vehicle 1100 itself must decide whether to heed the results from the primary computer or the secondary computer (e.g., the first controller 1136 or the second controller 1136). For example, in some embodiments, the ADAS system 1138 can be a secondary and / or auxiliary computer for providing perception information to a backup computer plausibility module. The backup computer plausibility monitor can run redundant diverse software on hardware components to detect faults in perception and dynamic driving tasks. The output from the ADAS system 1138 can be provided to a supervisory MCU. If the outputs from the primary and secondary computers conflict, the supervisory MCU must determine how to reconcile the conflict to ensure safe operation.

[0190] In some examples, the host computer can be configured to provide a confidence score to the supervisory MCU indicating the host computer's confidence in the selected result. If the confidence score exceeds a threshold, then the supervisory MCU can follow the host computer's direction, regardless of whether the secondary computer provides conflicting or inconsistent results. In the event that the confidence score does not satisfy the threshold and in the event that the host computer and the secondary computer indicate different results (e.g., a conflict), the supervisory MCU can arbitrate between the computers to determine the appropriate result.

[0191] The supervisory MCU can be configured to run a neural network that is trained and configured to determine conditions under which the secondary computer provides false alarms based at least in part on the output from the host computer and the secondary computer. Thus, the neural network in the supervisory MCU can learn when the output of the secondary computer can be trusted and when it cannot. For example, when the secondary computer is a RADAR-based FCW system, the neural network in the supervisory MCU can learn when the FCW system is identifying metal objects that are not in fact dangerous, such as drain grates or manhole covers that trigger false alarms. Similarly, when the secondary computer is a camera-based LDW system, the neural network in the supervisory MCU can learn to disregard the LDW when a cyclist or pedestrian is present and lane departure is in fact the safest strategy. In embodiments that include a neural network running on the supervisory MCU, the supervisory MCU can include at least one of a DLA or a GPU suitable for running a neural network with associated memory. In preferred embodiments, the supervisory MCU can include and / or be included as a component of the SoC 1104.

[0192] In other examples, the ADAS system 1138 can include a secondary computer that performs ADAS functions using traditional computer vision rules. As such, the secondary 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, the diverse implementation and intentional non-identity make the overall system more fault-tolerant, especially with respect to faults caused by software (or software-hardware interface) functions. For example, if there is a software bug or error in the software running on the host computer and the non-identical software code running on the secondary computer provides the same overall result, then the supervisory MCU can be more confident that the overall result is correct and that the bug in the software or hardware on the host computer did not cause a substantial error.

[0193] In some examples, the output of the ADAS system 1138 can be fed to a perception block of the host computer and / or a dynamic driving task block of the host computer. For example, if the ADAS system 1138 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 secondary computer can have its own neural network that is trained and thus reduces the risk of false positives as described herein.

[0194] The vehicle 1100 can also include an infotainment SoC 1130 (e.g., an in-vehicle infotainment system (IVI)). Although illustrated and described as a SoC, the infotainment system can not be a SoC and can include two or more discrete components. The infotainment SoC 1130 can include a combination of hardware and software that can be used to provide audio (e.g., music, a personal digital assistant, navigation instructions, news, radio, etc.), video (e.g., TV, movies, streaming media, etc.), telephony (e.g., hands-free calling), network connectivity (e.g., LTE, WiFi, etc.), and / or information services (e.g., a navigation system, a park assist, a radio data system, vehicle-related information such as a fuel level, a total distance covered, a brake fuel level, an oil level, a door open / close, air filter information, etc.) to the vehicle 1100. For example, the infotainment SoC 1130 can include a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, an in-car computer, in-car entertainment, WiFi, steering wheel audio controls, hands-free voice controls, a heads-up display (HUD), an HMI display 1134, 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 1130 can further be used to provide information (e.g., visual and / or audible) to a user of the vehicle, such as information from the ADAS system 1138, 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.

[0195] The infotainment SoC 1130 can include GPU functionality. The infotainment SoC 1130 can communicate with other devices, systems, and / or components of the vehicle 1100 over a bus 1102 (e.g., a CAN bus, Ethernet, etc.). In some examples, the infotainment SoC 1130 can be coupled to a supervisory MCU such that, in the event of a failure of the host controller 1136 (e.g., a primary and / or backup computer of the vehicle 1100), the GPU of the infotainment system can perform some autonomous driving functions. In such examples, the infotainment SoC 1130 can place the vehicle 1100 in a driver safe park mode as described herein.

[0196] The vehicle 1100 can also include an instrument cluster 1132 (e.g., a digital dashboard, an electronic instrument cluster, a digital instrument panel, etc.). The instrument cluster 1132 can include a controller and / or supercomputer (e.g., a discrete controller or supercomputer). The instrument cluster 1132 can include a set of instruments, such as a speedometer, fuel level, oil pressure, tachometer, odometer, turn indicator, shift position indicator, seat belt warning light, parking brake warning light, engine malfunction light, supplemental restraint system (SRS) system information, lighting controls, safety system controls, navigation information, and so forth. In some examples, information can be displayed and / or shared between the infotainment SoC 1130 and the instrument cluster 1132. In other words, the instrument cluster 1132 can be included as part of the infotainment SoC 1130, or vice versa.

[0197] Figure 11D FIG. 11 illustrates a system diagram of communication between a cloud-based server and an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. Figure 11A FIG. 11 illustrates a system diagram of communication between a cloud-based server and an example autonomous vehicle 1100 in accordance with some embodiments of the present disclosure. The system 1176 can include servers 1178, a network 1190, and vehicles including the vehicle 1100. The servers 1178 can include a plurality of GPUs 1184(A)- 1184(H) (collectively referred to herein as GPUs 1184), PCIe switches 1182(A)- 1182(H) (collectively referred to herein as PCIe switches 1182), and / or CPUs 1180(A)- 1180(B) (collectively referred to herein as CPUs 1180). The GPUs 1184, CPUs 1180, and PCIe switches can be interconnected with high-speed interconnects such as, for example and without limitation, NVLink interfaces 1188 developed by NVIDIA and / or PCIe connections 1186. In some examples, the GPUs 1184 are connected via NVLink and / or NVSwitch SoC connections, and the GPUs 1184 and PCIe switches 1182 are connected via PCIe interconnects. Although eight GPUs 1184, two CPUs 1180, and two PCIe switches are illustrated, this is not intended to be limiting. Depending on the embodiment, each of the servers 1178 can include any number of GPUs 1184, CPUs 1180, and / or PCIe switches. For example, each of the servers 1178 can include eight, sixteen, thirty-two, and / or more GPUs 1184.

[0198] The server 1178 can receive image data from vehicles over the network 1190 and representing images showing unexpected or changing road conditions such as a road work that has recently started. The server 1178 can transmit neural networks 1192, updated neural networks 1192, and / or map information 1194, including information about traffic and road conditions, to vehicles over the network 1190. Updates to the map information 1194 can include updates to the HD map 1122, e.g., information about construction sites, potholes, curves, flooding, or other obstacles. In some examples, the neural networks 1192, updated neural networks 1192, and / or map information 1194 can have been produced from experience using training performed at a data center (e.g., using the server 1178 and / or other servers) and / or from data received from any number of vehicles in the environment.

[0199] The server 1178 can be used to train machine learning models (e.g., neural networks) based on training data. The training data can be generated by vehicles and / or can be generated in simulations (e.g., using game engines). In some examples, the training data is labeled (e.g., in cases where the neural network benefits from supervised learning) and / or undergoes other pre-processing, while in other examples, the training data is not labeled and / or pre-processed (e.g., in cases where the neural network does not require supervised learning). The 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, federated learning, transfer learning, feature learning (including principal component and cluster analysis), multilinear subspace learning, manifold learning, representation learning (including spare dictionary learning), rule-based machine learning, anomaly detection, and any variants or combinations thereof. Once the machine learning models are trained, the machine learning models can be used by vehicles (e.g., transmitted to vehicles over the network 1190) and / or the machine learning models can be used by the server 1178 to remotely monitor vehicles.

[0200] In some examples, the server 1178 can receive data from vehicles and apply the data to the latest real-time neural networks for real-time intelligent inference. The server 1178 can include deep learning supercomputers and / or specialized AI computers powered by GPUs 1184, such as the DGX and DGX Station machines developed by NVIDIA. However, in some examples, the server 1178 can include deep learning infrastructure of a data center that is powered by CPUs only.

[0201] The deep learning infrastructure of the server 1178 can be capable of fast real-time inference, and can use this capability to assess and validate the health of the processors, software, and / or associated hardware in the vehicle 1100. For example, the deep learning infrastructure can receive periodic updates from the vehicle 1100, such as a sequence of images and / or objects located in the sequence of images that the vehicle 1100 has located (e.g., via computer vision and / or other machine learning object classification techniques). The deep learning infrastructure can run its own neural network to identify the objects and compare them to the objects identified by the vehicle 1100, and if the results do not match and the infrastructure concludes that the AI in the vehicle 1100 is malfunctioning, the server 1178 can transmit a signal to the vehicle 1100 instructing the fail-safe computer of the vehicle 1100 to take control, notify the passengers, and complete a safe parking operation.

[0202] For inference, the server 1178 can include GPUs 1184 and one or more programmable inference accelerators (such as NVIDIA’s TensorRT). The combination of GPU-powered servers and inference-accelerated can make real-time response possible. In other examples, such as where performance is less important, CPU, FPGA, and other processor-powered servers can be used for inference.

[0203] Example Computing Device

[0204] Figure 12 is a block diagram of an example computing device 1200 suitable for implementing some embodiments of the present disclosure. The computing device 1200 can include an interconnection system 1202 that directly or indirectly couples the following devices: memory 1204, one or more central processing units (CPUs) 1206, one or more graphics processing units (GPUs) 1208, a communication interface 1210, input / output (I / O) ports 1212, I / O components 1214, a power supply 1216, one or more presentation components 1218 (e.g., display(s)), and one or more logic units 1220. In at least one embodiment, the computing device(s) 1200 can include one or more virtual machines (VMs), and / or any component thereof can include a virtual component (e.g., a virtual hardware component). For a non-limiting example, one or more of the GPUs 1208 can include one or more vGPUs, one or more of the CPUs 1206 can include one or more vCPUs, and / or one or more of the logic units 1220 can include one or more virtual logic units. As such, the computing device(s) 1200 can include discrete components (e.g., a full GPU dedicated to the computing device 1200), virtual components (e.g., a portion of a GPU dedicated to the computing device 1200), or a combination thereof.

[0205] Although Figure 12 various blocks of the computing device are shown as being connected by the interconnect system 1202, this is not intended to be limiting and is for clarity only. For example, in some embodiments, a presentation component 1218, such as a display device, can be considered an I / O component 1214 (e.g., if the display is a touch screen). As another example, the CPU 1206 and / or GPU 1208 can include memory (e.g., the memory 1204 can represent a storage device in addition to the memory of the GPU 1208, CPU 1206, and / or other components). In other words, Figure 12 the computing device of FIG. 12 is illustrative. No distinction is made between Figure 12 such categories as “workstation,” “server,” “laptop,” “desktop,” “tablet,” “client device,” “mobile device,” “handheld device,” “game console,” “electronic control unit (ECU),” “virtual reality system,” and / or other device or system types, as all are considered within the scope of the computing device of FIG. 12.

[0206] The interconnect system 1202 can represent one or more links or buses, such as an address bus, data bus, control bus, or a combination thereof. The interconnect system 1202 can include one or more bus or link 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 or link. In some embodiments, there are direct connections between components. As an example, the CPU 1206 can be directly connected to the memory 1204. Further, the CPU 1206 can be directly connected to the GPU 1208. Where there are direct or point-to-point connections between components, the interconnect system 1202 can include a PCIe link to perform the connection. In these examples, a PCI bus need not be included in the computing device 1200.

[0207] The memory 1204 can include any of a variety of computer-readable media. Computer-readable media can be any available media that can be accessed by the computing device 1200. Computer-readable media can include both volatile and nonvolatile media, and removable and non-removable media. By way of example, and not limitation, computer-readable media can comprise computer storage media and communication media.

[0208] Computer storage media can include volatile and nonvolatile media and / or removable and non-removable media implemented in any method or technology for storage of information such as computer readable instructions, data structures, program modules, and / or other data types. For example, memory 1204 can store computer readable instructions (e.g., representing program(s) and / or program element(s), such as an operating system). Computer storage media can include, but is not limited to, RAM, ROM, EEPROM, flash memory or other memory technology, CD-ROM, digital versatile disks (DVD) or other optical disk storage, magnetic cassettes, magnetic tape, magnetic disk storage or other magnetic storage devices, or any other medium which can be used to store the desired information and which can be accessed by the computing device 1200. As used herein, computer storage media does not include signals per se.

[0209] Computer storage media can embody computer readable instructions, data structures, program modules, and / or other data types in a modulated data signal, such as a carrier wave or other transport mechanism, and includes any information delivery media. The term "modulated data signal" can refer to a signal that has one or more of its characteristics set or changed in such a manner as to encode information in the signal. By way of example, and not limitation, computer storage media can include wired media such as a wired network or direct-wired connection, and wireless media such as acoustic, RF, infrared, and other wireless media. Combinations of the any of the above should also be included within the scope of computer readable media.

[0210] The CPUs 1206 can be configured to execute at least some of the computer readable instructions to control one or more components of the computing device 1200 to perform one or more of the methods and / or processes described herein. The CPUs 1206 can each include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of handling a number of software threads concurrently. The CPUs 1206 can include any type of processors and can include different types of processors depending on the type of computing device 1200 being implemented (e.g., a mobile device having fewer cores for a processor and a server having more cores for a processor). For example, depending on the type of computing device 1200, the processor can be an Advanced RISC Machines (ARM) processor implemented using a Reduced Instruction Set Computing (RISC) or an x86 processor implemented using Complex Instruction Set Computing (CISC). The computing device 1200 can include one or more CPUs 1206 in addition to, or as an alternative to, one or more microprocessors or co-processors such as a mathematical co-processor.

[0211] In addition to, or in place of, CPU(s) 1206, GPU(s) 1208 can be configured to execute at least some of the computer-readable instructions to control one or more components of computing device 1200 to perform one or more of the methods and / or processes described herein. One or more of GPUs 1208 can be integrated GPUs (e.g., with one or more of CPUs 1206) and / or one or more of GPUs 1208 can be discrete GPUs. In embodiments, one or more of GPUs 1208 can be a co-processor of one or more of CPUs 1206. GPUs 1208 can be used by computing device 1200 to render graphics (e.g., 3D graphics) or to perform general purpose computing. For example, GPUs 1208 can be used for general purpose computing on GPUs (GPGPU). GPUs 1208 can include hundreds or thousands of cores capable of handling hundreds or thousands of software threads concurrently. GPUs 1208 can generate pixel data for output images in response to rendering commands (e.g., received from CPUs 1206 via a host interface). GPUs 1208 can include graphics memory (e.g., display memory) for storing pixel data or any other suitable data (e.g., GPGPU data). The display memory can be included as part of memory 1204. GPUs 1208 can include two or more GPUs operating in parallel (e.g., via a link). The link can connect the GPUs directly (e.g., using NVLINK) or can connect the GPUs through a switch (e.g., using an NVSwitch). When combined together, each GPU 1208 can generate pixel data or GPGPU data for a different portion of an output or for a different output (e.g., a first GPU for a first image and a second GPU for a second image). Each GPU can include its own memory, or can share memory with other GPUs.

[0212] In addition to or in place of CPU(s) 1206 and / or GPU(s) 1208, logic unit(s) 1220 can be configured to execute at least some of the computer-readable instructions to control one or more components of computing device 1200 to perform one or more of the methods and / or processes described herein. In embodiments, CPU(s) 1206, GPU(s) 1208, and / or logic unit(s) 1220 can perform any combination of the methods, processes, and / or portions thereof, discretely or jointly. One or more of logic unit(s) 1220 can be part of and / or integrated with one or more of CPU(s) 1206 and / or GPU(s) 1208, and / or one or more of logic unit(s) 1220 can be discrete components or otherwise external to CPU(s) 1206 and / or GPU(s) 1208. In embodiments, one or more of logic unit(s) 1220 can be a co-processor of one or more of CPU(s) 1206 and / or GPU(s) 1208.

[0213] Examples of logic unit(s) 1220 include one or more processing cores and / or components thereof, such as tensor cores (TCs), tensor processing units (TPUs), pixel visual cores (PVCs), visual processing units (VPUs), graphics processing clusters (GPCs), texture processing clusters (TPCs), streaming multi-processors (SMs), tree traversal units (TTUs), artificial intelligence accelerators (AIAs), deep learning accelerators (DLAs), arithmetic logic units (ALUs), application-specific integrated circuits (ASICs), floating point units (FPUs), input / output (I / O) elements, peripheral component interconnects (PCI) or peripheral component interconnect express (PCIe) elements, and the like.

[0214] Communication interface 1210 can include one or more receivers, transmitters, and / or transceivers that enable computing device 1200 to communicate with other computing devices via electronic communication networks, including wired and / or wireless communications. Communication interface 1210 can include components and functionality to enable communication over any of a plurality of different networks, such as wireless networks (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, etc.), wired networks (e.g., through Ethernet or InfiniBand), low power wide area networks (e.g., LoRaWAN, SigFox, etc.), and / or the Internet.

[0215] The I / O port(s) 1212 can enable the computing device 1200 to logically couple to other devices including I / O components 1214, presentation components 1218, and / or other components, some of which can be built into (e.g., integrated in) the computing device 1200. Illustrative I / O components 1214 include a microphone, mouse, keyboard, joystick, game pad, game controller, satellite dish, scanner, printer, wireless device, etc. The I / O components 1214 can provide a natural user interface (NUI) that processes air gestures, voice, or other biological inputs generated by a user. In some cases, inputs are transmitted to an appropriate network element for further processing. A NUI can implement any combination of speech recognition, gesture recognition, facial recognition, biometric recognition, posture recognition, gesture recognition within, as well as on the surface of, a display, air gestures, head and eye tracking, and touch recognition (as described in more detail below) associated with a display on the computing device 1200. The computing device 1200 can include a depth camera, such as a stereoscopic camera system, an infrared camera system, a RGB camera system, touchscreen technology, and combinations of these, for gesture detection and recognition. In addition, the computing device 1200 can include an accelerometer or a gyroscope (e.g., as part of an inertial measurement unit (IMU)) to detect motion. In some examples, the computing device 1200 can use the output of the accelerometer or gyroscope to render an immersive augmented reality or virtual reality.

[0216] The power supply 1216 can include a hard-wired power supply, a battery power supply, or a combination thereof. The power supply 1216 can provide power to the computing device 1200 to enable the components of the computing device 1200 to operate.

[0217] The presentation components 1218 can include a display (e.g., a monitor, a touch screen, a television, a heads-up display (HUD), other display types, or combinations thereof), speakers, and / or other presentation components. The presentation components 1218 can receive data from other components (e.g., the GPU 1208, the CPU 1206, etc.) and output the data (e.g., as a

[0218] Example data center

[0219] Figure 13 An example data center 1300 that can be used in at least one embodiment of the present disclosure is shown. The data center 1300 can include a data center infrastructure layer 1310, a framework layer 1320, a software layer 1330, and / or an application layer 1340.

[0220] As Figure 13As shown, the data center infrastructure layer 1310 can include a resource orchestrator 1312, grouped computing resources 1314, and node computing resources (“node C.R.s”) 1316(1)-1316(N), where “N” represents any whole, positive integer. In at least one embodiment, the node C.R.s 1316(1)-1316(N) can include, but are not limited to, any number of central processing units (“CPUs” or “processors”) including accelerators, field programmable gate arrays (FPGAs), graphics processors or graphics processing units (GPUs), memory devices (e.g., dynamic random access memory), storage devices (e.g., solid state or disk drives), network input / output (“NW I / O”) devices, network switches, virtual machines (“VMs”), power modules, and / or cooling modules, and the like. In some embodiments, one or more node C.R.s from the node C.R.s 1316(1)-1316(N) can correspond to a server having one or more of the above-described computing resources. Moreover, in some embodiments, the node C.R.s 1316(1)-1316(N) can include one or more virtual components, such as a vGPU, a vCPU, etc., and / or one or more of the node C.R.s 1316(1)-1316(N) can correspond to a virtual machine (VM).

[0221] In at least one embodiment, the grouped computing resources 1314 can include individual groups of node C.R.s 1316 housed within one or more racks (not shown) or housed within many racks at different geographic locations (also not shown). The individual groups of node C.R.s 1316 within the grouped computing resources 1314 can include grouped computing, network, memory, or storage resources that can be configured or allocated to support one or more workloads. In at least one embodiment, several node C.R.s 1316 including CPUs, GPUs, and / or other processors can be grouped within one or more racks to provide computing resources to support one or more workloads. The one or more racks can also include any quantity of power modules, cooling modules, and / or network switches in any combination.

[0222] The resource orchestrator 1322 can configure or otherwise control the one or more node C.R.s 1316(1)-1316(N) and / or the grouped computing resources 1314. In at least one embodiment, the resource orchestrator 1322 can comprise a software design infrastructure (“SDI”) management entity for the data center 1300. The resource orchestrator 1322 can comprise hardware, software, or some combination thereof.

[0223] In at least one embodiment, as Figure 13As shown, framework layer 1320 may include a job scheduler 1333, a configuration manager 1334, a resource manager 1336, and / or a distributed file system 1338. Framework layer 1320 may include a framework for software 1332 supporting software layer 1330 and / or one or more applications 1342 supporting application layer 1340. Software 1332 or application 1342 may respectively contain web-based service software or applications, such as those provided by Amazon Web Services, Google Cloud, and Microsoft Azure. Framework layer 1320 may be, but is not limited to, free and open-source software web application frameworks (such as Apache Spark) that can utilize distributed file system 1338 for large-scale data processing (e.g., "big data"). TM (Hereinafter referred to as "Spark") is a type of resource manager. In at least one embodiment, the job scheduler 1333 may include Spark drivers to facilitate the scheduling of workloads supported by different layers of data center 1300. The configuration manager 1334 may be able to configure different layers, such as the software layer 1330 and the framework layer 1320 (which includes Spark and a distributed file system 1338 for supporting large-scale data processing). The resource manager 1336 may be able to manage compute resources mapped to or allocated to clusters of distributed file system 1338 and job scheduler 1333 or to support clusters of distributed file system 1338 and job scheduler 1333. In at least one embodiment, the clustered or grouped compute resources may include grouped compute resources 1314 in the data center infrastructure layer 1310. The resource manager 1036 may coordinate with the resource coordinator 1312 to manage these mapped or allocated compute resources.

[0224] In at least one embodiment, the software 1332 included in software layer 1330 may include software used in at least a portion of the nodes CRs 1316(1)-1316(N), the grouped computing resources 1314, and / or the distributed file system 1338 of framework layer 1320. One or more types of software may include, but are not limited to, internet web search software, email virus scanning software, database software, and streaming video content software.

[0225] In at least one embodiment, applications 1342 included in application layer 1340 can include one or more types of applications used by at least portions of node C.R.s 1316(1)-1316(N), grouped computing resources 1314, and / or distributed file system 1338 of framework layer 1320. One or more types of applications can include, but are not limited to, any number of genomics applications, cognitive computing and machine learning applications including training or inference software, machine learning framework software (e.g., PyTorch, TensorFlow, Caffe, etc.), and / or other machine learning applications used in conjunction with one or more embodiments.

[0226] In at least one embodiment, any of configuration manager 1334, resource manager 1336, and resource orchestrator 1312 can implement any number and type of self-modification actions based on any quantity and type of data acquired in any technically feasible manner. Self-modification actions can free data center operators of data center 1300 from making potentially poor configuration decisions and can avoid underutilization and / or poor performance portions of data center.

[0227] According to one or more embodiments described herein, data center 1300 can include tools, services, software, or other resources to train one or more machine learning models or use one or more machine learning models to predict or infer information. For example, machine learning model(s) can be trained by computing weight parameters according to a neural network architecture using software and / or computing resources described above with respect to data center 1300. In at least one embodiment, trained or deployed machine learning models corresponding to one or more neural networks can be used to infer or predict information using resources described above with respect to data center 1300 by using weight parameters computed through one or more training techniques such as, but not limited to, those described herein.

[0228] In at least one embodiment, data center 1300 can use CPUs, application specific integrated circuits (ASICs), GPUs, FPGAs, and / or other hardware (or virtual computing resources corresponding thereto) to perform training and / or inference using resources described above. Further, one or more software and / or hardware resources described above can be configured as a service that allows users to train or perform inference on information, such as image recognition, speech recognition, or other artificial intelligence services.

[0229] Example network environment

[0230] Network environments suitable for implementing embodiments of the present disclosure can include one or more client devices, servers, network-attached storage (NAS), other backend devices, and / or other device types. The client devices, servers, and / or other device types (e.g., each device) can be implemented on one or more instances of the computing device(s) 1200 - e.g., each device can include similar components, features, and / or functionality of the computing device(s) 1200. Further, where backend devices (e.g., servers, NAS, etc.) are implemented, the backend devices can be included as part of a data center 1300, an example of which is described in more detail herein with respect to FIG. 2, and which can be implemented in accordance with the computing device(s) 1200. Figure 12 Figure 13

[0231] Components of the network environment can communicate with each other via a network, which can be wired, wireless, or both. The network can include multiple networks or one of multiple networks. For example, the network can include one or more wide area networks (WANs), one or more local area networks (LANs), one or more public networks (such as the Internet and / or the public switched telephone network (PSTN)), and / or one or more private networks. Where the network includes a wireless telecommunication network, components such as base stations, communication towers, or even access points (among other components) can provide wireless connectivity.

[0232] Compatible network environments can include one or more peer-to-peer network environments (in which case servers can not be included in the network environment) and one or more client-server network environments (in which case one or more servers can be included in the network environment). In a peer-to-peer network environment, functionality described herein for servers can be implemented on any number of client devices.

[0233] In at least one embodiment, the network environment can include one or more cloud-based network environments, distributed computing environments, combinations thereof, and the like. A cloud-based network environment can include a framework layer, a job scheduler, a resource manager, and a distributed file system implemented on one or more servers, which can include one or more core network servers and / or edge servers. The framework layer can include a framework that supports one or more applications of a software layer and / or an application layer. The software or applications can include network-based service software or applications, respectively. In embodiments, one or more client devices can use the network-based service software or applications (e.g., by accessing the service software and / or applications via one or more application programming interfaces (APIs)). The framework layer can be, without limitation, a free and open-source software web application framework as can be used for large-scale data processing (e.g., “big data”) using a distributed file system. ​​

[0234] The cloud-based network environment can provide cloud computing and / or cloud storage that performs any combination of the computing and / or data storage functions described herein (or one or more portions thereof). Any of these different functions can be distributed across multiple locations from a central or core server (e.g., can be distributed across one or more data centers in a state, region, country, globally, etc.). The core server can designate at least a portion of the functions to an edge server if the connection to the user (e.g., client device) is relatively close to the edge server. The cloud-based network environment can be private (e.g., limited to a single organization), public (e.g., available to many organizations), and / or a combination thereof (e.g., a hybrid cloud environment).

[0235] The client device(s) can include at least some of the components, features, and functionality of the example computing device 1200 described herein with respect to Figure 12 As examples and not by way of limitation, a client device can be implemented as a personal computer (PC), laptop computer, mobile device, smartphone, tablet computer, smartwatch, wearable computer, personal digital assistant (PDA), MP3 player, virtual reality headset, global positioning system (GPS) or device, video player, camera, surveillance device or system, vehicle, boat, spaceship, virtual machine, drone, robot, handheld communication device, hospital device, gaming device or system, entertainment system, vehicle computer system, embedded system controller, remote control, appliance, consumer electronic device, workstation, edge device, any combination of these depicted devices, or any other suitable device.

[0236] The present disclosure can be described in the general context of machine-usable instructions or computer code, including computer-executable instructions such as program modules, being executed by a computer or other machine, such as a personal data assistant or other handheld device. Generally, program modules include routines, programs, objects, components, data structures, etc., that perform particular tasks or implement particular abstract data types. The present disclosure can be practiced in a variety of system configurations, including hand-held devices, consumer electronics, general- purpose computers, more specialty computing devices, etc. The present disclosure can also be practiced in distributed computing environments where tasks are performed by remote processing devices that are linked through a communications network.

[0237] As used herein, the term "and / or," with respect to a listing of two or more items, means that one, including only one of the items, or combinations of the items, can be employed. For example, "A and / or B" can include A alone, B alone, or A and B. Likewise, "at least one of A or B" can include A alone, B alone, or at least one of A and at least one of B. Further, "at least one of A and B" can include at least one of A, at least one of B, or at least one of A and at least one of B.

[0238] The subject matter of the present disclosure is described with specificity herein to meet statutory requirements. However, the description itself is not intended to limit the scope of this disclosure. Rather, the present disclosure has been presented for the purpose of illustration and description so as to enable others, skilled in the art, to employ the subject matter of the present disclosure in various ways. Furthermore, although the terms "step" and / or "block" can be used herein to connote different elements of methods employed, the terms should not be interpreted as implying any particular order among or between various steps herein disclosed unless and except the order of individual steps is explicitly described.

Claims

1. A method for landmark detection, comprising: Apply sensor data representing the sensor's sensing field to neural networks; Using the neural network and at least partially based on the sensor data, calculate one or more positions of one or more control points of one or more curves identified in an image corresponding to the sensor data, and classification information of the one or more curves; use a curve fitting algorithm and at least partially based on the one or more positions of the one or more control points to generate the one or more curves; associate the classification information with the one or more curves; and have a machine perform one or more operations based at least partially on the one or more curves and the classification information.

2. The method of claim 1, wherein calculating the one or more positions of one or more control points comprises: The neural network is used to compute first data, which indicates the number of pixels in the image within the radius of at least one of the one or more curves identified in the image; Calculate second data, which indicates the distance to one or more control points of the one or more curves; and decode the first data and the second data to determine the one or more positions of the one or more control points.

3. The method according to claim 1, wherein the classification information represents one or more bit classification masks.

4. The method of claim 3, wherein the number of bit classification masks in the one or more bit classification masks is equal to one less than the number of classification types for which the neural network computes its output.

5. The method of claim 2, wherein the first data represents a binary mask, wherein the pixels of the image within the radius of at least one of the one or more curves depicted in the image have a first value, and the other pixels have a second value different from the first value.

6. The method of claim 2, wherein each curve comprises more than three control points, the second data represents multiple regression masks, and the multiple regression masks comprise multiple x-axis distance regression masks and y-axis distance regression masks equal in number to the control points.

7. The method according to claim 1, wherein the curve fitting algorithm includes at least one of a two-dimensional Bézier curve fitting algorithm, a three-dimensional Bézier curve fitting algorithm, a cubic Bézier curve fitting algorithm, a higher-order Bézier curve fitting algorithm, or a split Bézier curve fitting algorithm.

8. The method according to claim 1, further comprising: A clustering algorithm is performed on one or more curves to generate a final set of curves, wherein the one or more operations are performed based at least in part on the final set of curves and the classification information.

9. The method of claim 8, wherein the clustering algorithm includes a density-based spatial clustering DBSCAN operation with noise, the DBSCAN operation using an integral loss as a similarity function.

10. The method of claim 1, wherein the one or more curves correspond to one or more of a road marking, lane line, road boundary line, intersection line, sidewalk, bicycle lane line, text, pole, tree, lamppost or sign.

11. A system for landmark detection, comprising: One or more sensors; One or more processing units; And one or more memory devices storing instructions thereon, which, when the instructions are executed using the one or more processing units, cause the one or more processing units to perform the following operations, including: applying sensor data representing an image corresponding to the sensing field of the sensor in the one or more sensors to the neural network; Using the neural network and at least partially based on the sensor data, calculate one or more positions of one or more control points of one or more curves corresponding to one or more features identified in the image, and classification information of the one or more curves; use a curve fitting algorithm and at least partially based on the one or more positions of the one or more control points to generate the one or more curves; associate the classification information with the one or more curves; and perform one or more operations based at least partially on the one or more curves and the classification information.

12. The system of claim 11, wherein the calculation comprises: Calculate a binary mask indicating the pixels of the image within the radius of at least one of one or more curves corresponding to features identified in the image; calculate one or more distance regression masks indicating the distances to one or more control points of the one or more curves; and decode the binary mask and the distance regression mask to determine the one or more locations of the one or more control points.

13. The system of claim 12, wherein the calculation of the classification information of the one or more curves comprises using the neural network and at least in part based on the sensor data to calculate one or more bit classification masks indicating the classification information of the one or more curves.

14. The system of claim 13, wherein the number of bit classification masks in the one or more bit classification masks is equal to one less than the number of classification types for which the neural network computes its output.

15. The system of claim 12, wherein each curve includes a plurality of control points, and the distance regression mask includes a plurality of x-axis distance regression masks and y-axis distance regression masks equal in number to the number of control points.

16. The system according to claim 11, wherein the curve fitting algorithm includes at least one of a two-dimensional Bézier curve fitting algorithm, a three-dimensional Bézier curve fitting algorithm, a cubic Bézier curve fitting algorithm, a higher-order Bézier curve fitting algorithm, or a split Bézier curve fitting algorithm.

17. The system of claim 11, wherein the operation further comprises: A clustering algorithm is performed on one or more curves to generate a final set of curves, wherein the one or more operations are performed based at least in part on the final set of curves and the classification information.

18. The system of claim 11, wherein the system comprises at least one of the following: a system for performing simulation operations; a system for performing simulation operations to test or validate autonomous machine applications; a system for performing deep learning operations; a system implemented using edge devices; a system for merging one or more virtual machines (VMs); a system implemented at least partially in a data center; or a system implemented at least partially using cloud computing resources.

19. A method for training a neural network, comprising: Receive annotation data representing polylines associated with an image and classification information associated with the polylines; Based at least in part on the polyline, a curve fitting algorithm is performed to determine multiple control points associated with the curve; based at least in part on the multiple control points associated with the curve, a binary mask is encoded for the pixels of the image indicating the radius of the curve; And encode multiple distance regression masks, each distance regression mask indicating the distance to a control point among the multiple control points; At least in part based on the classification information, one or more bit classification masks are encoded; The neural network is trained using the binary mask, the plurality of distance regression masks, and the one or more bit classification masks as ground reality data.

20. The method of claim 19, further comprising: At least one of the top or bottom portions of the polyline is trimmed to generate a trimmed polyline, wherein the curve fitting algorithm is performed at least in part based on the trimmed polyline.

21. The method of claim 19, further comprising: Receive initial annotation data representing the polygon; And at least in part based on the initial annotation data, the polyline is extracted as one of the left edge, right edge, or center rail of the polygon to generate the annotation data.

22. The method of claim 19, wherein the neural network is trained using one or more of integral loss, binary cross-entropy loss, masked L1 loss, or masked binary cross-entropy loss.

23. The method of claim 19, wherein encoding the plurality of distance regression masks comprises: For each of the plurality of control points, the x-axis distance regression mask and the y-axis distance regression mask are encoded.

24. The method of claim 19, wherein the neural network is trained to predict a number n of classification types, and encoding the one or more bit classification masks comprises: Encode the n-1 bit classification mask.

25. The method according to claim 19, wherein the curve fitting algorithm includes at least one of a two-dimensional Bézier curve fitting algorithm, a three-dimensional Bézier curve fitting algorithm, a cubic Bézier curve fitting algorithm, a higher-order Bézier curve fitting algorithm, or a split Bézier curve fitting algorithm.

Images