Projecting images captured using fisheye lenses for feature detection in autonomous machine applications

Abstract

In an example, features in an environment of a vehicle can be detected with live perception from a wide field of view sensor. Sensor data generated by the sensor can be adjusted to represent a virtual field of view that is different from the actual field of view of the sensor, and the sensor data, with or without the virtual adjustment, can be applied to a stereo projection algorithm to generate a projected image. The projected image can then be applied to a machine learning model, such as a deep neural network (DNN), to detect and / or classify features or objects represented therein. The machine learning model can be pre-trained on training sensor data generated by a sensor having a field of view that is less than the field of view of the wide field of view sensor, such that the virtual adjustment and / or the projection algorithm can update the sensor data to be suitable for accurate processing by the pre-trained machine learning model.

Description

BACKGROUND

[0001] Autonomous driving systems and advanced driver assistance systems (ADAS) can utilize various sensors to perform various tasks—such as lane keeping, lane changing, lane assignment, camera calibration, steering, stopping, path planning, and localization. For example, for autonomous and ADAS systems to operate independently and effectively, a real-time or near real-time understanding of the vehicle’s surroundings can be generated. This understanding can include information about the location of objects, obstacles, lane markings, signs, and / or traffic lights in the environment that provide context and visual markers for various divisions such as lanes, road boundaries, intersections, etc. The vehicle can use the information of the surroundings when making decisions such as given the travel paths of various objects (e.g., vehicles, pedestrians, bicyclists, etc.) in the environment, when and whether to change lanes, how fast to drive, where to stop at an intersection, etc.

[0002] As an example, information about the location and properties of objects and / or lanes in the environment of an autonomous or semi-autonomous vehicle can prove valuable when performing path planning, obstacle avoidance, and / or control decisions. Machine learning models and / or computer vision algorithms are often trained or programmed to generate information of a vehicle’s surroundings. For example, these machine learning models (e.g., deep neural networks (DNNs)) and / or computer vision algorithms are trained to generate an understanding of the surroundings represented by sensor data (e.g., images) generated by sensors with varying fields of view. For example, many DNNs can be trained using sensors (e.g., image sensors) with fields of view between 60 degrees and 120 degrees. However, at least some sensors of a vehicle (e.g., cameras, LIDAR sensors, RADAR sensors, etc.) can have a field of view greater than 120 degrees, such as a parking camera positioned at the back or front of the vehicle, or a side view camera positioned on a side mirror of the vehicle. For example, parking cameras often employ fisheye cameras with a field of view upwards of or greater than 190 degrees. Thus, image data generated by these image sensors with wide fields of view can not be suitable for processing by a DNN—e.g., due to distortion, artifacts, and / or other imperfections.

[0003] In conventional systems, objects of a vehicle’s environment can be detected using a DNN trained to detect features represented by training image data generated using image sensors having a field of view less than 120 degrees. For example, pinhole cameras, commonly used for front-facing cameras (colloquially known as “dash cams” or “drive recorders”) and rear-facing cameras, are examples of image sensors having a generally narrow field of view. Images captured by such image sensors have minimal distortion, as these images are mostly straight lines, and images generated by such cameras can be used to train a DNN to detect features of the environment. However, image data captured by parking cameras (e.g., fisheye cameras) having a higher field of view (e.g., greater than 120 degrees) can include distortion in areas where limited information is available (e.g., the edges of the image), and thus can result in inaccurate calculations by the DNN. As such, conventional approaches can retrain the DNN using image data captured by sensors having a larger field of view, which not only requires significant computational cost and manual effort, but also limits the scalability and adaptability of the DNN to represent sensor data of smaller fields of view (e.g., less than 120 degrees). Thus, these conventional systems require various instances of the DNN to be trained, such that each instance of the DNN corresponds to a particular field of view— e.g., a first instance for a field of view of 90-120 degrees and a second instance for a field of view of 120-180 degrees. Furthermore, even in cases where the DNN is specifically trained for a wider field of view, due to the variation in both the proportion and angle of distortion in image data from wide field of view sensors, it can be difficult to ground truth label features due to the tilted orientation of features (e.g., objects, lines) in the image. For example, features such as lines and shapes in image-space can not align with the location of the same features in world-space, adding another layer of complexity for the DNN and / or post-processing to accurately reconcile the output of the DNN with the real-world location of the features. Without accurately mapping the output to world-space locations, the DNN can not be as reliably used to perform operations in the technical space as safely critical as autonomous driving. SUMMARY

[0004] Embodiments of the present disclosure relate to stereophotogrammetrically projecting images captured using fisheye lenses for feature detection using neural networks. Systems and methods are disclosed that leverage existing neural networks trained on outputs captured using narrower field of view sensors (e.g., sensors having a field of view less than 120 degrees) to detect features in outputs from wider field of view sensors (e.g., sensors having a field of view greater than 120 degrees) in real-time or near real-time.

[0005] In contrast to conventional systems (such as those described above), the systems and methods of the present disclosure can utilize live perception of a wide field of view sensor (e.g., greater than 120 degrees) to detect one or more features in the vehicle’s environment. For example, images from the wide field of view sensor can be applied to a stereographic projection algorithm to project the images onto a two-dimensional (2D) plane. The projected images can then be utilized to detect features in the vehicle’s environment using a neural network trained to detect features in images captured by a narrower field of view sensor. In some examples, the field of view of the wide field of view sensor can be virtually adjusted to generate updated images having a virtually adjusted field of view prior to applying the images to the stereographic projection algorithm. The field of view of the wide field of view sensor can be adjusted vertically such that a virtual center (e.g., camera center point) of the sensor is substantially aligned with the horizon. In some other examples, detected features can be converted to image-space locations and corresponding world-space locations. For example, the output can be used to directly or indirectly (e.g., via decoding) determine a location of each feature, a classification of each feature, and / or the like.

[0006] As a result of using an existing neural network (e.g., a DNN trained using lower field of view images) to detect features in the output of a high field of view image, additional computational and time resources are not required to train a new neural network or retrain a pre-trained neural network for a wide field of view image sensor. As such, the process of detecting features in images captured using a wide field of view sensor can be relatively less time consuming, less computationally intensive, and more scalable, as the system can learn to detect features in real-time or near real-time without requiring prior experience, training, or knowledge of the environment and field of view of the sensor. BRIEF DESCRIPTION OF DRAWINGS

[0007] The systems and methods for stereographically projecting images captured using fisheye lenses for feature detection of the present disclosure are described in detail below with reference to the attached drawing figures, wherein:

[0008] Figure 1A is an example data flow diagram illustrating an example process for detecting features of a vehicle’s environment using output from one or more sensors of the vehicle in accordance with some embodiments of the present disclosure;

[0009] Figure 1B is an illustration depicting an example of projecting an original image onto a 2D projection plane using a virtual sphere to generate a projected image in accordance with some embodiments of the present disclosure;

[0010] Figure 2 is an illustration depicting an example field of view of a wide field of view sensor on a vehicle in accordance with some embodiments of the present disclosure;

[0011] Figure 3FIG. 1 is a diagram depicting example distortion representations corresponding to three sensors with different fields of view, in accordance with some embodiments of the present disclosure;

[0012] Figure 4 FIG. 2 is a diagram depicting example images captured with virtual adjustment of field of view, in accordance with some embodiments of the present disclosure;

[0013] Figure 5 FIG. 3 is a diagram depicting example object detection in a projected image, in accordance with some embodiments of the present disclosure;

[0014] Figure 6 FIG. 4 is a flowchart showing an example process for detecting features in an image captured by a wide field of view sensor using an existing neural network trained on images captured using a narrower field of view sensor, in accordance with some embodiments of the present disclosure;

[0015] Figure 7 FIG. 5 is a flowchart showing an example process for detecting features in a virtually adjusted projected image captured by a wide field of view sensor, in accordance with some embodiments of the present disclosure;

[0016] Figure 8A FIG. 6 is a diagram of an example autonomous vehicle, in accordance with some embodiments of the present disclosure;

[0017] Figure 8B FIG. 7 is an example of a camera position and field of view of the example autonomous vehicle of Figure 8A FIG. 8 is an example of a camera position and field of view of the example autonomous vehicle of

[0018] Figure 8C FIG. 9 is an example of a camera position and field of view of the example autonomous vehicle of Figure 8A FIG. 10 is a block diagram of an example system architecture of the example autonomous vehicle of

[0019] Figure 8D FIG. 11 is a system diagram for communication between a cloud-based server and the example autonomous vehicle of Figure 8A FIG. 12 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure.

[0020] Figure 9 FIG. 13 is a block diagram of an example computing device suitable for implementing some embodiments of the present disclosure. DETAILED DESCRIPTION

[0021] Systems and methods are disclosed that relate to using neural networks to stereographically project images captured using fisheye lenses for feature detection. While the systems and methods can be described with respect to an example autonomous vehicle 800 (or, in this document, referred to as “vehicle 800” or “own vehicle 800”), the systems and methods can be used with respect to any vehicle having a fisheye lens, such as a vehicle having a fisheye lens for a rearview mirror. Figures 8A-8DThe present disclosure is described with reference to examples) described herein, but this is not intended to be limiting. For example, the systems and methods described herein can be used by, but are not limited to, non-autonomous vehicles, semi-autonomous vehicles (e.g., in one or more adaptive driver assistance systems (ADAS)), robots, warehouse vehicles, off-road vehicles, flying watercraft, boats, shuttles, emergency response vehicles, motorcycles, electric or motorized bicycles, airplanes, construction vehicles, underwater vessels, drones, and / or other vehicle types. Further, while the present disclosure can be described with reference to feature detection and classification for vehicle applications, this is not intended to be limiting and the systems and methods described herein can be used in augmented reality, virtual reality, robotics, security and surveillance, autonomous or semi-autonomous machine applications, and / or any other technical space in which wide field of view sensors are deployed and can be used to perform various operations (e.g., object and / or feature detection or classification).

[0022] As described herein, the current systems and methods provide techniques for using existing deep neural networks (DNNs) trained on images captured using narrower field of view sensors (e.g., sensors having a field of view of less than 120 degrees) to detect features in images captured by wide field of view sensors (e.g., sensors having a field of view of greater than 120 degrees) in real-time or near real-time, as compared to conventional approaches. As such, features (e.g., objects, lanes, pedestrians) in images of a vehicle’s environment can be detected with live perception from a wide field of view sensor. In some embodiments, image data from a wide field of view sensor can be projected onto a two-dimensional target plane to generate a stereographic projected image with less distortion (e.g., on the edges of the image) that less information is available. After stereographic projection, the image data can be applied to an existing DNN trained on sensor data representing a narrower field of view, such that the DNN does not need to be retrained for a wide field of view image sensor. However, in other embodiments, a pre-trained DNN can be omitted by training a DNN on the projected and / or virtually adjusted image (e.g., an image with a virtually adjusted field of view) without departing from the scope of the present disclosure.

[0023] In some embodiments, to accurately project highly distorted regions captured in image data, a field of view of an image represented by image data can be virtually adjusted prior to applying the image data to a stereographic algorithm. For example, a field of view of an image sensor can be adjusted upward (e.g., because a parking image sensor or side view image sensor can be angled downward and toward a ground plane). In some embodiments, the field of view can be adjusted (e.g., rotated) so that a virtual center of the image sensor is substantially aligned with a horizontal line (e.g., a horizontal line in the real world). For example, an amount of rotation can be determined based on an analysis of a mounting angle of the image sensor, an analysis of individual images captured using the image sensor, and / or other information. As such, a predetermined adjustment angle can be determined for the virtually adjusted field of view. In some embodiments, this adjustment can correspond to generally aligning a point on the image sensor (e.g., a camera center) with a horizontal line in the environment. In some embodiments, the horizontal line can correspond to a boundary delineating a driving surface from a sky in the image. In such embodiments, the adjustment to the field of view can include, for one or more calibration images, adjusting the field of view until a point on the image sensor is substantially aligned with a horizontal line as represented in the one or more calibration images. Once this alignment is determined, the adjustment angle can be determined (e.g., by averaging the angle over any number of calibration images), and the determined adjustment angle can be used as a virtual adjustment to the field of view. By virtually adjusting the field of view, pixels in a highly distorted region of an image in which valuable information is represented can become less distorted, and a portion of the image data in which more information is available (e.g., a center of the image) can be more distorted - but even after the distortion, the amount of information available at the center of the image can be sufficient for accurate processing.

[0024] Additionally, in some embodiments, feature detection can be used to determine locations of features and / or objects in world-space. In such examples, sensor calibration (e.g., intrinsic and / or extrinsic properties of the sensor) can be used to convert image-space locations to world-space locations. The world-space locations of features and / or objects can then be used by various other systems of the vehicle to perform path planning, obstacle avoidance, control decisions, and / or other autonomous or semi-autonomous operations.

[0025] As such, live perception by a high field of view camera can be leveraged to generate an understanding of a vehicle’s environment using existing DNNs - e.g., without retraining the DNNs or training new DNNs. As such, the amount of computational power and manpower required to perform object detection, feature detection, and / or other computations using sensor data from a wide field of view sensor can be significantly reduced compared to conventional systems. Moreover, even where DNNs are trained on projected and / or virtually adjusted field of view images, these results can be more accurate and reliable compared to conventional systems that attempt to train DNNs on unprojected and / or unadjusted fields of view.

[0026] Object and feature detection system

[0027] At a high level, sensors (e.g., cameras, RADAR sensors, LIDAR sensors, etc.) located or otherwise arranged on autonomous or semi-autonomous vehicles can be used to receive and / or generate sensor data (e.g., image data, LIDAR data, RADAR data, etc.). The sensors can be wide-field-of-view sensors (e.g., sensors with a field of view equal to or greater than 120 degrees). The sensor data can be applied to projection algorithms trained and / or programmed to generate projected sensor data representing projected images (e.g., stereo projection algorithms, sphere-centric projection algorithms, etc.). In the case of using a stereo projection algorithm, the algorithm can project pixels of the sensor data onto a projection plane based on a ray formula on a virtual sphere depicting the field of view of the wide-field-of-view sensor. Projected sensor data can be applied to neural networks (e.g., deep neural networks (DNNs), such as convolutional neural networks (CNNs)) trained to recognize regions of interest belonging to (as a non-limiting example) objects in an environment (e.g., vehicles, pedestrians, etc.), features of the environment (e.g., raised pavement markings, vibration strips, colored lane dividers, sidewalks, intersections, forks, road layouts, objects, etc.), and / or semantic information associated with them (e.g., waiting conditions, object types, lane types). In some examples, the image-spatial location of detected objects or features can also be computed by the DNN. As described herein, the DNN can be trained (e.g., pre-trained) using images captured from narrow field-of-view sensors (e.g., sensors with a field of view of less than 120 degrees) or other sensor data representations.

[0028] In some embodiments, spherical projection algorithms can be used to perform projection of sensor data to generate projected sensor data for use by the DNN. However, spherical projection may be less accurate for wide field-of-view sensors because it can fail or be unreliable for fields of view greater than 90 degrees, where the projection of some pixels in the image data may reach infinity on the target plane. Accordingly, to address this issue, a subset of the image data can be projected onto different planes to address a wider field of view. However, such partial projection can be computationally expensive, as at least four planes may be required to cover the entire field of view. Furthermore, features may end up in portions on different projection planes, leading to inaccurate DNN outputs.

[0029] Accordingly, in some embodiments, a stereographic algorithm can be used to project the pixels of an image onto a target plane (e.g., a two-dimensional plane). A virtual sphere can be used as a virtual field of view of the sensor, and the lowest point on the virtual sphere can be used as a center of projection to project the image onto the target plane. For each pixel on the target plane (e.g., the projected image), based on the intersection of a virtual line between the center of projection and the pixel on the target plane and a point (e.g., a pixel) on the virtual sphere, it can be determined that the point on the sphere is projected onto that pixel. In this way, each pixel on the target plane or projected image corresponds to a pixel that is sampled on the original image. By producing the projected image in this way, each original image is fully captured on a two-dimensional plane, such that the projected image is invertible, where features detected on the projected image by a neural network can be traced back to the original image using the ray formula to determine the location of the features on the original image. Further, the stereographic projection algorithm can project images captured by a wide field of view sensor onto a single plane, thereby relatively reducing the computational cost as compared to epi- polar techniques that divide an image into sections and then project the sections on multiple planes. The projected image can include a planar view of the original image captured by a sensor that has a field of view between 120 degrees and 360 degrees by preserving regions of interest in the original unprojected image that have the most distortion.

[0030] In some examples, the sensor data can undergo pre-processing to virtually adjust the field of view of the wide field of view sensor to generate an updated image, such as regions (e.g., edges) of the image that are most distorted - and potentially informative - can be relocated to regions of the virtual sphere where pixel information is most preserved during projection. For example, the virtual field of view of the wide field of view sensor can be adjusted in the vertical direction such that the virtual center of the sensor is substantially aligned with the horizon in the real world. In such examples, the virtual center of the image can be moved vertically upward by rotating the virtual sphere by a predetermined number of degrees. The predetermined number of degrees can be based on sensor calibration, and as a non-limiting example, can be between 20 degrees and 60 degrees. For example, the number of degrees of rotation can be determined based on the location and angle of the sensor on the vehicle. In some examples, the virtual adjustment can be produced by rotating the rays that form the original image by the predetermined number of degrees when projecting the image onto the projected plane. The updated image can be applied to the stereographic projection algorithm to generate the projected image for application to the DNN. Aligning the virtual center of the sensor with the horizon can allow for a constant alignment for all sensor data generated by the sensor, and can allow for the edges of the image represented by the sensor data to be aligned with the center position on the virtual sphere for preserving the most information.

[0031] The projected images can be applied to a DNN (e.g., a pre-trained DNN) to detect objects, features, and / or semantic information corresponding thereto. In embodiments, the output of the DNN can be used to accurately track objects as feature detection can be preserved at the edges of the images captured by the wide field of view sensor - e.g., portions of the image that can be the most distorted and thus most difficult to predict without stereophotography.

[0032] Once features and / or objects are detected, their corresponding locations can be converted to their respective world-space locations. This can be done using calibration information corresponding to the sensors and can be based on adjustments (e.g., vertical rotation) to the sensor data during processing. Thus, the original mapping of image-space locations to world-space locations from unprocessed sensor data can be recovered in order to prepare the output of the DNN for use by the vehicle to perform one or more operations.

[0033] Reference Figure 1A , Figure 1A is an example dataflow diagram illustrating an example process 100 for detecting features of an environment of a vehicle using output from one or more sensors of the vehicle, 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 and that components and / or processes can be adjusted without departing from the scope of the present disclosure. Further, additional or alternative components and / or processes can be implemented in addition to or instead of those described herein.

[0034] The process 100 can include a stereo projector 106 that performs a stereo projection algorithm on one or more inputs (e.g., image data 102, other sensor data, etc.) and generates one or more outputs - such as projected image data 108. Further, the process 100 can include one or more machine learning models 110 (e.g., DNNs) that receive one or more inputs (e.g., projected image data 108, with or without field of view adjustments from the FOV adjuster 104) and generate one or more outputs 112. In some examples, the image data 102 can be referred to as training image data when used for training. Although the image data 102 is primarily discussed with respect to image data representing images, this is not intended to be limiting and the image data 102 can include other types of sensor data for feature and / or object detection, such as, for example, LIDAR data, SONAR data, RADAR data, etc., generated by one or more sensors of the vehicle 800( Figures 8A-8D ) as described herein.

[0035] Process 100 can include generating and / or receiving image data 102 from one or more sensors. As non-limiting examples, image data 102 can be received from one or more sensors of a vehicle (e.g., vehicle 800 described herein) captured by a wide field of view sensor (e.g., a sensor having a field of view greater than 120 degrees) Figures 8A-8C and used by the vehicle, and within process 100, to detect and / or classify objects or features to help navigate through the vehicle’s environment in real-time or near real-time. Image data 102 can include, but is not limited to, image data 102 from any sensor of a vehicle including, for example and with reference to Figures 8A-8C , one or more stereo cameras 868, one or more wide-angle cameras 870 (e.g., fisheye cameras), one or more infrared cameras 872, one or more surround cameras 874 (e.g., 360 degree cameras), and / or one or more long and / or mid-range cameras. In some embodiments, in addition to or alternatively to image data 102, sensor data from any number of sensor types can be used, such as but not limited to RADAR sensors 860, ultrasonic sensors 862, LIDAR sensors 864, and / or other sensor types. In some embodiments, as described herein, image data 102 and / or sensor data can be generated with sensors having a field of view or field of view greater than 120 degrees (e.g., 120 to 360 degrees). As another example, image data 102 can include virtual image data generated from any number of sensors of a virtual vehicle or other virtual object. In such examples, the virtual sensors can correspond to a virtual vehicle or other virtual object in a simulated environment (e.g., to test, train, and / or validate neural network performance), and the virtual image data can represent image data captured by the virtual sensors within the simulated or virtual environment.

[0036] In some embodiments, image data 102 can include image data representing images, image data representing video (e.g., a snapshot of a video), and / or sensor data representing a representation of a sensory field of a sensor captured by a wide field of view sensor (e.g., a sensor having a field of view greater than 120 degrees) (e.g., a depth map of a LIDAR sensor, a value map of an ultrasonic sensor, etc.). With respect to image data 102, any type of image data format can be used, such as, but not limited to, compressed images (such as Joint Photographic Experts Group (JPEG) or luminance / chrominance (YUV) formats), compressed images (such as frames derived from a compressed video format (such as H.264 / Advanced Video Coding (AVC) or H.265 / High Efficiency Video Coding (HEVC)), raw images (such as derived from a Red Clear Blue (RCCB), Red Clear Clear (RCCC), or other type of imaging sensor), and / or other formats. Additionally, in some instances, image data 102 can be used within process 100 without any pre-processing (e.g., in a raw or captured format), while in other examples, image data 102 can undergo pre-processing (e.g., any one or more of noise balancing, demosaicing, scaling, cropping, augmentation, white balancing, tone curve adjustment, etc., such as using a sensor data pre-processor (not shown)). As used herein, image data 102 can refer to unprocessed image data, pre-processed image data, or a combination thereof.

[0037] Image data 102 can include raw images (e.g., as captured by one or more image sensors), down-sampled images, up-sampled images, cropped or region of interest (ROI) images, otherwise augmented images, and / or combinations thereof. In some embodiments, one or more machine learning models 110 can be trained using images (and / or other image data 102) captured using a narrower field of view sensor (e.g., a sensor having a field of view less than 120 degrees).

[0038] As an example of a field of view greater than 120 degrees, and with reference to Figure 2 , Figure 2 A field of view of a wide field of view sensor on a vehicle is shown in accordance with some embodiments of the disclosure. For example, vehicle 210 includes two wide field of view sensors, sensor 212 and sensor 214. Field of view 222 can correspond to sensor 212 and can be approximately 190 degrees. Similarly, field of view 224 can correspond to sensor 214 and can be approximately 190 degrees. Sensor 212 and sensor 214 can each represent a fisheye camera of vehicle 210 (e.g., wide-angle camera 870 of vehicle 800). In some examples, one or both of sensor 212 and sensor 214 can be a parking camera used to assist vehicle 210 in parking, or can include a side view camera used to assist vehicle 210 in lane changing, blind spot monitoring, etc.

[0039] Referring again to Figure 1A During deployment, the image data 102 can be applied to the FOV adjuster 104, for example, to compute updated image data to be applied to the stereographic projector 106. In some instances, the image data 102 can undergo pre-processing to virtually adjust the field of view of the wide field sensor to produce updated image data such that the most distorted and potentially informative and important regions of the image (e.g., the edges) can be relocated to regions of the virtual sphere where pixel information is most preserved during projection. For example, an image captured using a narrower field of view sensor can have minimal distortion near the edges of the field of view (e.g., as shown in image 310 of FIG. 3). However, an image captured using a higher field of view sensor typically has heavy distortion at the edges, where less information is available for each pixel (e.g., as shown in image 330 of FIG. 3). Figure 3 Figure 3

[0040] For example, the FOV adjuster 104 can adjust the virtual field of view of the wide field sensor in the vertical direction (e.g., up or down). Due to the mounting angle of the sensor (e.g., facing down for a parking sensor and side view sensor, facing up for an intersection analysis sensor, etc.), the field of view of the sensor can not be ideal for each embodiment. For a non-limiting example, a parking sensor angled 40 degrees toward the ground can not have an ideal field of view for detecting vehicles in an adjacent lane or parked along a side of the street. However, it can still be advantageous to utilize the image data generated by the parking sensor angled downward for object or feature detection tasks. Thus, the image data can be updated to represent a virtual field of view of the parking sensor angled less than 40 degrees toward the ground (e.g., to 20 degrees or 0 degrees). In some embodiments, as described herein, the field of view can be virtually adjusted such that the virtual center of the sensor is substantially aligned with the horizon in the real world. In such examples, the virtual center of the image in the image data 102 can be moved vertically, for example, up or down, by rotating the virtual sphere by a predetermined number of degrees. The virtual sphere can serve as the virtual field of view of the wide field sensor that captured the image data 102, and the updated image data representing the virtually adjusted field of view can be applied to the stereographic projector 106.

[0041] ​​In some examples, the predetermined degree can be based on sensor calibration, and as a non-limiting example, can be between 20 degrees to 60 degrees. For example, the FOV adjuster 104 can determine a degree of rotation of the virtual sphere based on the position and angle of the sensor on the vehicle 800. In some examples, the virtual adjustment can be generated by rotating the rays of the original image that form the image data 102 by the predetermined degree to generate an updated image. Aligning the virtual center of the wide field of view sensor with the horizon can allow edges of the image in the image data 102 to be aligned with the center position on the virtual sphere for preserving information that is most distorted in the unadjusted image.

[0042] As another example, and with reference to Figure 3 , Figure 3 images captured using three sensors with different fields of view are shown in accordance with some embodiments of the disclosure. Image 310 is an image captured using a sensor with a horizontal field of view of sixty degrees. As can be seen in image 310, there is substantially no distortion in the image. Most machine learning models (e.g., one or more machine learning models 110) are trained to detect features using training images with a similar field of view as image 310. Due to the lack of distortion, the machine learning models can be able to accurately and efficiently detect features in images such as image 310. Similarly, image 320 is an image captured using a sensor with a horizontal field of view of approximately 120 degrees. As the field of view increases, the distortion at the edges of the captured image also increases. As can be seen, image 320 includes distortion on the right and left edges of the image. However, image 320 can be almost entirely or substantially straight, and the machine learning models can still be able to accurately detect features throughout image 320, as the loss of information in image 320 is still minimal. Image 330 is an image captured using a sensor with a field of view of 190 degrees. The pixels at the left and right edges of the image exhibit a higher level of distortion than the pixels in the middle of image 330. In cases where a machine learning model such as one or more machine learning models 110 is trained using images captured by a narrower field of view sensor (e.g., a sensor with a field of view of less than 120 degrees), the machine learning model can not be able to detect features in the distorted regions, such as the edges of image 330, without pre-processing. Thus, process 100 can be used to pre-process images captured using a field of view closer to the field of view represented in image 330, such that machine learning models 110 (e.g., pre-trained on lower field of view images and / or trained on adjusted field of view or projected images) can make accurate predictions with respect to image 330.

[0043] As a further example, and with respect to Figure 4 , Figure 4A virtually adjusted image adjusted using the FOV adjuster 104 is shown in accordance with some embodiments of the present disclosure. The image 410 can be an original image captured by a high field of view (e.g., greater than 120 degrees) sensor 440 of a vehicle (e.g., vehicle 800). Figure 1A The FOV adjuster 104 can be used to virtually adjust the field of view 444A of the wide field of view sensor 440 that captured the image 410 to generate an updated image such that the most distorted and potentially informative regions (e.g., edges 412 and 414) of the image can be relocated to regions of the virtual sphere where pixel information is most preserved during projection (e.g., regions 422 and 424). The updated image 420 can be generated by vertically adjusting the field of view 444A of the sensor 440 to an updated field of view 444B such that the virtual center of the sensor is virtually aligned with regions in the real world that are different from the actual alignment of the sensor in the real world (e.g., such as the horizon 426 in the real world). As can be seen, the regions 422 and 424 of the updated image 420 preserve pixel information from the edges 412 and 414 of the image 410. For example, the image 410 can be captured by the wide field of view sensor 440 such that the center of the sensor is facing down towards the street or ground plane, and after the adjustment in the image 420 the virtual center of the sensor is more facing the horizon 426 (e.g., the intersection of the road or ground plane with the sky in the distance). As such, by rotating the virtual center of the sensor upwards, the regions 422 and 424 that are the most distorted in the image 410 can be updated such that pixel information is preserved during projection using the stereographic projector 106.

[0044] Referring again to Figure 1A The updated image data (e.g., after the field of view adjustment) and / or the image data 102 can be applied to the stereographic projector 106 trained and / or programmed to generate the projected image data 108. The stereographic projector 106 can perform a stereographic projection algorithm, a gnomonic projection algorithm, and / or another type of projection algorithm. The stereographic projector 106 can project the pixels of the image data 102 onto a two-dimensional (2D) projection plane (e.g., a target plane) based on a light ray formula on a virtual sphere that depicts the field of view of the wide field of view sensor used to capture the image data 102. The virtual sphere can serve as the field of view of the wide field of view sensor, and each pixel of the image data 102 can be projected onto the 2D projection plane. The projected image data 108 can represent a projected image.

[0045] In some examples, the stereographic projector 106 can use a spherical center projection algorithm (and can alternatively be referred to as a“spherical center projector” or“projector”) to generate the projected image data 108. The center of a virtual sphere can be used as the center of the projection to project the image data 102 onto a 2D projection plane. For each pixel of the 2D projection plane (e.g., the projected image data 108), based on the intersection of a virtual line between the center of the projection and the pixel on the target plane with a point (e.g., a pixel) on the virtual sphere, it can be determined that the point on the virtual sphere is projected onto that pixel. In this way, each pixel on the target plane or the projected image data 108 can correspond to a pixel that is sampled on the original image of the image data 102. However, the spherical center projection algorithm can not be able to project pixels in a field of view that is greater than 90 degrees, such that the remaining pixels of an image captured from a sensor having a field of view that is greater than 90 degrees are left outside of the 2D projection plane because the projection for those pixels of the image data 102 can reach infinity without intersecting the 2D projection plane. In such examples, a subset of the image data 102 can be projected onto separate 2D projection planes to account for a wider field of view (e.g., a field of view that is greater than 90 degrees). The separate 2D projection planes can be used as the projected image data 108 to be applied to the machine learning model 110 to detect features and / or objects. However, this partial and multiple projection can be computationally expensive because at least four projection planes can be needed to cover the entire field of view of the image data 102. Further, a single feature or object can end with a portion that is represented across different projection planes, making it difficult for the machine learning model 110 to predict such features and / or objects.

[0046] In other examples, the stereographic projector 106 can use a stereographic projection algorithm to generate the projected image data 108 by projecting the pixels of the image of the image data 102 or the updated image data onto a single 2D projection plane. In such examples, a virtual sphere can be used as the virtual field of view of the sensor, and the lowest (e.g., vertically lowest) point on the virtual sphere can be used as the center of the projection to project the image data 102 onto the 2D projection plane (e.g., as shown in FIG. 2B). In this way, the projected image data 108 can be generated by projecting the image data 102 onto a single 2D projection plane. However, the stereographic projection algorithm can not be able to project pixels in a field of view that is greater than 90 degrees, such that the remaining pixels of an image captured from a sensor having a field of view that is greater than 90 degrees are left outside of the 2D projection plane because the projection for those pixels of the image data 102 can reach infinity without intersecting the 2D projection plane. In such examples, a subset of the image data 102 can be projected onto separate 2D projection planes to account for a wider field of view (e.g., a field of view that is greater than 90 degrees). The separate 2D projection planes can be used as the projected image data 108 to be applied to the machine learning model 110 to detect features and / or objects. However, this partial and multiple projection can be computationally expensive because at least four projection planes can be needed to cover the entire field of view of the image data 102. Further, a single feature or object can end with a portion that is represented across different projection planes, making it difficult for the machine learning model 110 to predict such features and / or objects. Figure 1A and Figure 1BFor each pixel of the 2D projection plane, based on the intersection of the virtual line between the center of the projection and the pixel on the 2D projection plane and the point (e.g., pixel) on the virtual sphere, it can be determined that the point on the virtual sphere is projected onto that pixel. In this way, each pixel on the 2D projection plane or each pixel on the projected image of the projected image data 108 can correspond to a pixel that is sampled on the original image of the image data 102 and / or the updated image data. By generating the projected image in this way, each original image represented by the image data 102 and / or the updated image represented by the updated image data can be fully captured on a two-dimensional plane, such that the projected image is invertible, where the feature detected on the projected image by the machine learning model 110 can be traced back to the original image of the image data 102 using the ray formula to determine the location of the feature on the original image of the image data 102. As non-limiting examples, among others, Figure 1B In particular, it is shown how the original image is projected onto a 2D projection plane using a virtual sphere to generate the projected image data 108.

[0047] Further, the stereographic projector 106 can use a stereographic projection algorithm to project the images of the image data 102 captured by a wide field of view sensor onto a single plane, thereby relatively reducing the computational cost as compared to the gnomonic projection technique that divides the image into individual portions and then projects the divided portions on multiple planes. The projected image in the projected image data 102 can include a planar view of the original image of the image data 102 captured by a sensor having a field of view between 120 degrees and 360 degrees by preserving the region of interest in the original unprojected image that has the greatest distortion. In particular, the FOV adjuster 104 can move the distorted portion (e.g., edges) of the image data 102 to a location in the updated image data such that the distorted portion in the updated image data is in a region where the stereographic projector 106 is configured to preserve more information. While the non-distorted portion of the image data 102 can be in a region where the stereographic projector 106 can not be configured to preserve as much information, the image data 102 itself includes more information in those regions of the image (e.g., center) such that the machine learning model 110 can be able to accurately predict features in such regions even in the event of information loss. For example, referring to Figure 5 , the object (or vehicle) 514 can appear distorted after the projection, but the object 514 can still be accurately detected and / or classified by the machine learning model 110 due to the number of pixels corresponding to the object 514.

[0048] As an example of stereographic projection, and referring to Figure 1B , Figure 1BIt is shown how some embodiments according to the present disclosure use a virtual sphere to project an original image onto a 2D projection plane to generate a projected image. A virtual sphere 120 is used as a virtual field of view of a sensor that captures an image represented in the outline of the sphere, and the lowest (e.g., vertically lowest) point on the virtual sphere is used as a center 122 of the projection of the image onto a target plane 118. For each pixel of this target plane, a point on the virtual sphere is determined to be projected onto the pixel based on an intersection of a virtual line between the center 122 of the projection and the pixel on the target plane 118 with the point on the virtual sphere. For example, a point 128 on the virtual sphere 120 is determined to be projected onto a pixel 130 of the target plane 118 based on an intersection of a virtual line 126 between the center 122 and the pixel 130. Similarly, points 134, 140, 144, and 150 on the virtual sphere are projected onto pixels 136, 140, 146, and 152, respectively, of the target plane 118 based on intersections of virtual lines 132, 138, 142, and 148, respectively, between the center 122 and the respective pixels. As such, each pixel on the target plane 118 can correspond to a sampled pixel on the original image to generate a projected image.

[0049] Referring again to Figure 1A The projected image data 108 can be applied to a machine learning model 110 trained to detect an output 112 from the image data 102 — e.g., after the field of view adjustment and / or the projection. The machine learning model 110 can use the projected image data 108 to compute an output 112 that can be applied to a decoder or one or more post-processing components (e.g., output converters) to generate information about the environment of the vehicle 800. The machine learning model can be trained to identify regions of interest related to the environment of the vehicle 800 (e.g., as one of the outputs 112). For example, the regions of interest and the subsequent one or more outputs 112 can include objects (e.g., vehicles, pedestrians, stop signs, etc.), features (e.g., raised pavement markers, rumble strips, colored lane dividers, sidewalks, intersections, forks, road layouts, etc.), and / or semantic information (e.g., classifications, wait conditions, object types, lane types, etc.) related thereto. In some examples, the one or more machine learning models 110 can be further trained to determine image-space locations of the detected objects or features. In some embodiments, the one or more machine learning models 110 can be trained (e.g., pre-trained) with image or other sensor data representations captured from a narrower field of view sensor (e.g., a sensor having a field of view less than 120 degrees), as described herein.

[0050] Although examples are described herein with respect to using deep neural networks (DNNs) and, in particular, convolutional neural networks (CNNs) as machine learning models 110, this is not intended to be limiting. For example, and without limitation, one or more machine learning models 110 can include one or more machine learning models of any type, such as 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.), channel detection algorithms, computer vision algorithms, and / or other types of machine learning models.

[0051] As an example, such as where one or more machine learning models 110 include a CNN, one or more machine learning models 110 can include any number of layers. One or more of the layers or above can include an input layer. The input layer can hold values associated with image data 102 and / or projection image data 108 (e.g., before or after post-processing). For example, when image data 102 and / or projection image data 108 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).

[0052] One or more of the layers can include a convolutional layer. The convolutional layer can compute the outputs of neurons connected to local regions of the input layer, each neuron computing their weights with the dot product between their weights and the small region they are connected to in the input volume. The result of the convolutional layer can be another volume, where one dimension is based on the number of filters applied (e.g., width, height, and number of filters, such as 32x32x12, if 12 is the number of filters).

[0053] 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), which is thresholded at zero. The resulting volume of the ReLU layer can be the same as the volume of the input to the ReLU layer.

[0054] One or more of the layers can include a pooling layer. The pooling layer can perform a down-sampling operation along spatial dimensions (e.g., height and width), which can result in a smaller volume than the input to the pooling layer (e.g., 16x16x12 from a 32x32x12 input volume).

[0055] One or more of these layers may include one or more fully connected layers. Each neuron in a fully connected layer can be connected to every neuron in a previous volume. The fully connected layer can compute class scores, and the resulting volume can be 1x1x the number of classes. In some examples, a CNN may include fully connected layers such that the outputs of one or more layers of the CNN can be fed as input to the fully connected layers of the CNN. In some examples, one or more convolutional streams may be implemented by a machine learning model 110, and some or all of the convolutional streams may include corresponding fully connected layers.

[0056] In some non-limiting embodiments, the machine learning model 110 may include a series of convolutional layers and max pooling layers to facilitate image feature extraction, followed by multi-scale dilated convolutional layers and upsampling layers to facilitate global contextual feature extraction.

[0057] While the input layer, convolutional layer, pooling layer, ReLU layer, and fully connected layer are discussed in relation to machine learning model 110 in this paper, this is not intended to be restrictive. For example, additional or alternative layers, such as normalization layers, SoftMax layers, and / or other layer types, may be used in one or more machine learning models 110.

[0058] In embodiments where the machine learning model 110 includes a CNN, the order and number of layers of the CNN may vary depending on the embodiment. In other words, the order and number of layers in one or more machine learning models 110 are not limited to any one architecture.

[0059] Furthermore, some layers may include parameters (e.g., weights and / or biases), such as convolutional and fully connected layers, while others may not, such as ReLU layers and pooling layers. In some examples, the parameters may be learned by the machine learning model 110 during training. Further, some layers may include additional hyperparameters (e.g., learning rate, span, epoch, etc.), such as convolutional, fully connected, and pooling layers, while others may not, such as ReLU layers. The parameters and hyperparameters are not limited and may vary depending on the implementation.

[0060] As an example of output 112, and see... Figure 5 , Figure 5 According to some embodiments of this disclosure, object detection in an image using a neural network is illustrated. Image 500 may represent a vehicle (such as...) Figures 8A-8D Images captured by a wide field-of-view sensor (e.g., a sensor with a field of view greater than 120 degrees, a fisheye camera, a wide-angle camera 870, a surround camera 874, a parking assist camera, etc.) of the vehicle 800. As described herein, Figure 1AOne or more machine learning models 110 can be used to detect objects 512, 514, and 516 in the image 500. In applying the projected image data 108 to the machine learning models 110, the stereographic projector 106 can be used to project the image 500 onto a 2D target plane. In some non-limiting embodiments, the one or more machine learning models 110 can be trained using images captured from a narrower field of view sensor (e.g., a sensor having a field of view of less than 120 degrees). In this way, existing machine learning models can be leveraged to detect features in images captured from a different field of view than the training images without requiring retraining of the machine learning models 110. Moreover, due to the adjusted field of view and / or projection, features and / or objects can be detected in regions of the original image that were distorted prior to the field of view adjustment and / or projection - e.g., regions of the image corresponding to objects (e.g., vehicles) 512 and 516. As such, portions of the original image that can have the most distortion but represent important information (e.g., the location of objects 512 and 516) can be less distorted, and portions of the original image that are not distorted (e.g., corresponding to object 514) can have more distortion, but still maintain enough pixels to represent it such that the one or more machine learning models 110 can still accurately predict outputs 112 corresponding thereto. Figure 1B The stereographic projector 106 can be used to project the image 500 onto a 2D target plane prior to applying the projected image data 108 to the machine learning models 110. In some non-limiting embodiments, the one or more machine learning models 110 can be trained using images captured from a narrower field of view sensor (e.g., a sensor having a field of view of less than 120 degrees). In this way, existing machine learning models can be leveraged to detect features in images captured from a different field of view than the training images without requiring retraining of the machine learning models 110. Moreover, due to the adjusted field of view and / or projection, features and / or objects can be detected in regions of the original image that were distorted prior to the field of view adjustment and / or projection - e.g., regions of the image corresponding to objects (e.g., vehicles) 512 and 516. As such, portions of the original image that can have the most distortion but represent important information (e.g., the location of objects 512 and 516) can be less distorted, and portions of the original image that are not distorted (e.g., corresponding to object 514) can have more distortion, but still maintain enough pixels to represent it such that the one or more machine learning models 110 can still accurately predict outputs 112 corresponding thereto.

[0061] Referring again to Figure 1A In some examples, the one or more outputs 112 can be applied to an output converter 114 in order to post-process the one or more outputs 112 of the one or more machine learning models 110. In some examples, once a feature is detected in the one or more outputs 112 of the one or more machine learning models 110, a location corresponding to the feature can be converted or mapped to its respective world-space location. The output converter 114 can use calibration information (e.g., intrinsic and / or extrinsic parameters such as sensor model, position and orientation of the sensor on the vehicle 800, focal length, lens distortion, pose, etc.) corresponding to the sensor that captured the image data 102 and / or adjustments made to the image data 102 by the FOV adjuster 104 during processing (e.g., vertical rotation) to convert the image-space location from the output 112 to a world-space location. Thus, the original mapping of image-space locations from the unprocessed sensor data to world-space locations can be recovered in order to prepare the one or more outputs 112 of the one or more machine learning models 110 for use by the one or more control components 116 of the vehicle 800 in performing one or more operations.

[0062] In some embodiments, one or more outputs 112 of one or more machine learning models 110 can be used to accurately track objects. For example, because feature detection can be saved at the edges of images of image data 102 that are initially distorted by a wide field of view sensor, objects or features can be accurately tracked throughout the field of view (e.g., including portions of the image that can be most distorted and thus most difficult to make predictions about without stereographic projection). Outputs 112 containing features and / or objects in the environment can be tracked in subsequent images of image data 102, and in some instances, temporal analysis can be used to track features from when they are detected at one edge of the field of view to when they are detected at another edge of the field of view. As such, where tracking is performed, output converter 114 can use the real world converted outputs of objects and / or features to produce a history of locations or movements corresponding thereto.

[0063] Once the one or more outputs 112 are determined and / or converted, the information can be passed to one or more control components 116 of the system to perform one or more operations. For example, in the case where the system is the vehicle 800 described herein, the one or more outputs 112 and / or converted outputs can be passed to one or more layers of an autonomous driving software stack (e.g., a planning layer, a control layer, a world model manager, a perception layer, an obstacle avoidance layer of the driving stack, an actuation layer of the driving stack, etc.) to determine appropriate control decisions. Thus, the one or more control components 116 can make control decisions that can include one or more of suggested path planning, obstacle avoidance, and / or control decisions as to where to stop, how fast to drive, what path to use to safely traverse an environment, where other vehicles or pedestrians can be positioned, and / or the like. In any example, and with respect to autonomous or semi-autonomous driving, the control decisions can include any decisions corresponding to a perception layer of the driving stack, a world model management layer of the driving stack, a planning layer of the driving stack, a control layer of the driving stack, an obstacle avoidance layer of the driving stack, an actuation layer of the driving stack, and / or another layer, feature, or function of the driving stack. In some examples, the process 100 can be performed on any number of machine learning models 110 operating within the system. For example, an autonomous driving software stack can rely on hundreds or thousands of machine learning models 110 for effective and safe operation, and any number of these models can be subjected to the process 100 in order to ensure safe and effective operation while utilizing sensors with a large field of view (e.g., greater than 120 degrees). As such, the process 100 can be performed individually for any number of different operations corresponding to one or more layers of the driving stack and using any number of machine learning models 110 as described herein. As an example, a first detection can be determined for an object detection operation with respect to a perception layer of the driving stack using a first machine learning model, and a second detection can be determined for path planning with respect to a planning layer of the driving stack using a second machine learning model trained for regression on lane lines.

[0064] Referring now to Figure 6 and Figure 7 , each block of the methods 600 and 700 described herein comprises a computational process that can be performed using any combination of hardware, firmware, and / or software. For instance, different functions can be implemented by a processor executing instructions stored in memory. The methods 600 and 700 can also be embodied as computer-usable instructions stored on computer storage media. To this end, the methods 600 and 700 can be provided as a plurality of program instructions implemented on one or more computing devices. For example, the methods 600 and 700 can be provided as standalone software, as a service or hosted service (standalone or in combination with other hosted services), or as a plug-in to another product. Further, by way of example, the processes 100 and 300 of Figure 1A and Figures 8A-8DThe vehicle 800 describes the methods 600 and 700. However, these methods can additionally or alternatively be performed by any one system or within any one process, or any combination of systems and processes, including but not limited to those described herein.

[0065] Reference is now made to Figure 6 , Figure 6 is a flowchart illustrating a method 600 for detecting features in an image captured by a wide field of view sensor using an existing neural network trained on images captured using a narrower field of view sensor, according to some embodiments of the present disclosure. At block B602, the method 600 includes receiving image data representing an image generated using a first image sensor having a field of view. For example, the image data 102 can be received, where the image data 102 represents an image captured from a wide field of view sensor (e.g., a sensor having a field of view of 120 degrees or more).

[0066] At block B604, the method 600 includes applying the image data to a stereographic projection algorithm to generate projection image data representing a projected image. For example, the image data 102 can be applied to the stereographic projector 106 executing a stereographic projection algorithm to produce the projection image data 108 representing a projected image resulting from projecting the pixels of the image data 102 onto a 2D projection or target plane.

[0067] At block B606, the method 600 includes applying the projection image data to a neural network trained to detect features represented by training image data representing images generated using one or more second image sensors having a field of view smaller than the field of view. For example, the projection image data 108 can be applied to one or more machine learning models 110 trained to detect and / or classify features and / or objects represented by training image data representing images generated using a narrower field of view sensor (e.g., a sensor having a field of view smaller than 120 degrees).

[0068] At block B608, the method 600 includes computing data representing feature detections corresponding to one or more features using the neural network and based at least in part on the projection image data. For example, the one or more machine learning models 110 can compute one or more outputs 112 based on the projection image data 108. The outputs 112 can include data representing feature and / or object detections corresponding to one or more features in the projection image data 108.

[0069] Reference is now made to Figure 7 , Figure 7is a flowchart illustrating a method 700 for detecting features in an image acquired by a wide field of view sensor, in accordance with some embodiments of the present disclosure. At block B702, the method 700 includes receiving image data representing an image generated using a first image sensor having a field of view greater than or equal to 120 degrees. For example, the image data 102 can be received, where the image data represents an image captured from a wide field of view sensor (e.g., a sensor having a field of view of 120 degrees or more).

[0070] At block B704, the method 700 includes virtually adjusting the field of view of the first image sensor. For example, the FOV adjuster 104 can virtually adjust the field of view of the image sensor used to capture the image data 102.

[0071] At block B706, the method 700 includes generating updated image data corresponding to the virtually adjusted field of view based at least in part on the image data. For example, the updated imaging data can be generated by the FOV adjuster 104 that corresponds to the virtually adjusted field of view.

[0072] At block B708, the method 700 includes applying the updated image data to a stereographic projection algorithm to generate projection image data representing a projected image. For example, the updated image data can be applied to the stereographic projector 106 that performs a stereographic projection algorithm to produce the projection image data 108 representing a projected image produced by projecting the pixels of the updated image data onto a 2D projection or target plane.

[0073] At block B710, the method 700 includes applying the projection image data to one of a machine learning model or a computer vision algorithm. For example, the projection image data 108 can be applied to the machine learning model 110.

[0074] At block 712, the method 700 includes computing data representing feature detections corresponding to one or more features using the one of a machine learning model or a computer vision algorithm and based at least in part on the projection image data. For example, the one or more machine learning models 110 can compute one or more outputs 112 based on the projection image data 108. The outputs 112 can include data representing feature and / or object detections or classifications corresponding to one or more features and / or objects represented by the projection image data 108.

[0075] Example autonomous vehicle

[0076] Figure 8Ais a diagram of an example autonomous vehicle 800 according to some embodiments of the present disclosure. Autonomous vehicle 800 (alternatively referred to herein as “vehicle 800”) 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 drone and / or a vehicle 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 previous and future versions of this standard). Vehicle 800 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, vehicle 800 can be capable of implementing conditional automation (Level 3), high automation (Level 4), and / or full automation (Level 5).

[0077] Vehicle 800 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. Vehicle 800 can include a propulsion system 850 such as an internal combustion engine, a hybrid electric power plant, a fully electric motor, and / or another type of propulsion system. Propulsion system 850 can be connected to a drivetrain of vehicle 800 that can include a transmission in order to effect propulsion of vehicle 800. Propulsion system 850 can be controlled in response to receiving a signal from a throttle / accelerator 852.

[0078] A steering system 854, which can include a steering wheel, can be used to steer vehicle 800 (e.g., along a desired path or route) while propulsion system 850 is operating (e.g., while the vehicle is in motion). Steering system 854 can receive a signal from a steering actuator 856. For full automation (Level 5) functionality, a steering wheel can be optional.

[0079] A braking sensor system 846 can be used to operate vehicle brakes in response to receiving a signal from a braking actuator 848 and / or a braking sensor.

[0080] may include one or more system on chips (SoCs) 804 Figure 8C) and / or one or more controllers 836 of one or more GPUs can provide signals (e.g., representative of commands) to one or more components and / or systems of the vehicle 800. For example, the one or more controllers can send signals to operate vehicle brakes via one or more brake actuators 848, to operate a steering system 854 via one or more steering actuators 856, to operate a propulsion system 850 via one or more throttle / accelerator 852. The one or more controllers 836 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 representative of commands) to enable autonomous driving and / or to assist a human driver in driving the vehicle 800. The one or more controllers 836 can include a first controller 836 for autonomous driving functions, a second controller 836 for functional safety functions, a third controller 836 for artificial intelligence functions (e.g., computer vision), a fourth controller 836 for infotainment functions, a fifth controller 836 for redundancy in emergency situations, and / or other controllers. In some examples, a single controller 836 can handle two or more of the above functions, two or more controllers 836 can handle a single function, and / or any combination thereof.

[0081] The one or more controllers 836 can provide signals for controlling one or more components and / or systems of the vehicle 800 in response to sensor data (e.g., sensor inputs) received from one or more sensors. The sensor data can be received from, for example and without limitation, a global navigation satellite system sensor 858 (e.g., a global positioning system sensor), a RADAR sensor 860, an ultrasonic sensor 862, a LIDAR sensor 864, an inertial measurement unit (IMU) sensor 866 (e.g., an accelerometer, a gyroscope, a magnetic compass, a magnetometer, etc.), a microphone 896, a stereo camera 868, a wide-angle camera 870 (e.g., a fisheye camera), an infrared camera 872, a surround camera 874 (e.g., a 360-degree camera), a long-range and / or mid-range camera 898, a speed sensor 844 (e.g., to measure a speed of the vehicle 800), a vibration sensor 842, a steering sensor 840, a brake sensor (e.g., as part of a brake sensor system 846), and / or other sensor types.

[0082] One or more of the controllers 836 can receive input (e.g., represented by input data) from an instrument cluster 832 of the vehicle 800 and provide output (e.g., represented by output data, display data, etc.) via a human-machine interface (HMI) display 834, an audible annunciator, a speaker, and / or via other components of the vehicle 800. These outputs can include, for example and without limitation, vehicle speed, velocity, time, map data (e.g., represented by a map display), and / or other information. The one or more controllers 836 can also provide output to one or more of the vehicle’s systems (e.g., to control the vehicle’s systems) and / or to one or more of the vehicle’s components (e.g., to control the vehicle’s components). Figure 8Cinformation such as information about objects and object states as perceived by the controller 836, and so on. For example, the HMI display 834 can display information about the presence of one or more objects (e.g., a street sign, a warning sign, a traffic light change, and so on) 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, and so on).

[0083] The vehicle 800 also includes a network interface 824 that can communicate over one or more networks using one or more wireless antennas 826 and / or modems. For example, the network interface 824 can be capable of communicating over LTE, WCDMA, UMTS, GSM, CDMA2000, and so on. The one or more wireless antennas 826 can also enable communication between objects (e.g., vehicles, mobile devices, and so on) in the implementation environment using one or more local area networks such as Bluetooth, Bluetooth LE, Z-Wave, ZigBee, and so on and / or one or more low power wide area networks (LPWANs) such as LoRaWAN, SigFox, and so on.

[0084] Figure 8B Examples of camera positions and fields of view of the example autonomous vehicle 800 according to some embodiments of the present disclosure. Figure 8A Examples of camera positions and fields of view of the example autonomous vehicle 800 according to some embodiments of the present disclosure.

[0085] 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 800. 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), 820 fps, 240 fps, and so on, 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

[0086] In some examples, one or more of the cameras can be used to perform advanced driver assistance system (ADAS) functions (e.g., as part of a redundant or fail-safe design). For example, a multi-function monocular camera can be installed to provide functions including lane departure warning, traffic sign assist, and intelligent headlamp control. One or more of the cameras (e.g., all of the cameras) can simultaneously record and provide image data (e.g., video).

[0087] One or more of the cameras can be mounted in mounting assemblies such as custom designed (3-D printed) assemblies to cut off stray light and reflections from within the car that can interfere with the image data capture capabilities of the cameras (e.g., reflections from the dashboard reflected in the windshield mirror). With respect 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 mirror. In some examples, one or more cameras can be integrated into the wing mirror. For side view cameras, one or more cameras can also be integrated into the four pillars at each corner of the cab.

[0088] Cameras with fields of view that include the portion of the environment in front of the vehicle 800 (e.g., front-facing cameras) can be used for surround view to help identify the forward path and obstacles, and to assist in providing information critical to generating an occupancy grid and / or determining a preferred vehicle path with the help of one or more controllers 836 and / or control SoCs. Front-facing cameras can be used to perform many of the same ADAS functions as LIDAR, including emergency braking, pedestrian detection, and collision avoidance. Front-facing cameras can also be used for ADAS functions and systems including lane departure warning ("LDW"), adaptive cruise control ("ACC"), and / or other functions such as traffic sign recognition.

[0089] A wide variety of cameras can be used in the front-facing configuration, including, for example, monocular camera platforms including CMOS (complementary metal-oxide semiconductor) color imagers. Another example can be a wide-angle camera 870, which can be used to perceive objects (e.g., pedestrians, intersection traffic, or bicycles) entering the field of view from the periphery. Although Figure 8B Although only one wide-angle camera is illustrated in FIG. 8, there can be any number of wide-angle cameras 870 on the vehicle 800. In addition, long-range cameras 898 (e.g., pairs of stereo long-range cameras) can be used for depth-based object detection, especially for objects for which a neural network has not been trained. Long-range cameras 898 can also be used for object detection and classification and basic object tracking.

[0090] One or more stereo cameras 868 can also be included in the front-facing configuration. Stereo cameras 868 can include an integrated control unit that includes a scalable processing unit that can provide a multi-core microprocessor with integrated CAN or Ethernet interface and a field programmable logic (FPGA) on a single chip. Such a unit can be used to generate a 3-D map of the vehicle's environment, including distance estimates for all points in the image. Alternative stereo cameras 868 can include a compact stereo vision sensor that can include two camera lenses (one on the left and one on the right) and an image processing chip that can measure the distance from the vehicle to a target object and use the generated information (e.g., metadata) to activate autonomous emergency braking and lane departure warning functions. Other types of stereo cameras 868 can be used in addition to or instead of those described herein.

[0091] Cameras with fields of view that include portions of the environment to the side of vehicle 800 (e.g., side-view cameras) can be used for surround view, providing information used to create and update the occupancy grid and to generate side impact collision warnings. For example, surround cameras 874 (e.g., four surround cameras 874 as shown in FIG. 6B) can be placed on vehicle 800. Surround cameras 874 can include wide-view cameras 870, 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, a vehicle can use three surround cameras 874 (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. Figure 8B

[0092] Cameras with fields of view that include portions of the environment to the rear of vehicle 800 (e.g., rear-view cameras) can be used for assist parking, surround view, rear collision warnings, and to create and update the occupancy grid. A wide variety of cameras can be used, including but not limited to cameras that are also suitable as front-facing cameras as described herein (e.g., long- and / or mid-range cameras 898, stereo cameras 868, infrared cameras 872, etc.).

[0093] Figure 8C For use in a vehicle according to some embodiments of the present disclosure Figure 8A ​FIG. 1 is a block diagram of an example system architecture of an example autonomous vehicle 800. It should be understood that this arrangement and other arrangements described herein are set forth merely as examples. Other arrangements and elements (e.g., machines, interfaces, functions, orders, groupings of functions, etc.) can be used in addition to or instead of those shown, and some elements can be wholly omitted. 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 combinations and locations. Various functions described herein as being performed by an entity can be implemented in hardware, firmware, and / or software. For instance, various functions can be implemented by a processor executing instructions stored in a memory.

[0094] Figure 8C Each of the components, features, and systems of vehicle 800 are illustrated as being connected via bus 802. Bus 802 can include a controller area network (CAN) data interface (alternatively referred to herein as a "CAN bus"). The CAN can be a network within vehicle 800 that is used to assist in controlling various features and functions of vehicle 800, 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, revolutions per minute (RPM) of the engine, button positions, and / or other vehicle status indicators. The CAN bus can be ASIL B compliant.

[0095] Although bus 802 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 802 is represented with a single line, this is not intended to be limiting. For example, there can be any number of buses 802, 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 802 can be used to perform different functions, and / or can be used for redundancy. For example, a first bus 802 can be used for collision avoidance functions, and a second bus 802 can be used for drive control. In any example, each bus 802 can communicate with any component of vehicle 800, and two or more buses 802 can communicate with the same components. In some examples, each SoC 804, each controller 836, and / or each computer within the vehicle can have access to the same input data (e.g., inputs from sensors of vehicle 800), and can be connected to a common bus, such as a CAN bus.

[0096] Vehicle 800 may include one or more controllers 836, such as those described herein. Figure 8A The controllers described herein. Controller 836 can be used for a wide variety of functions. Controller 836 can be coupled to any other different components and systems of vehicle 800 and can be used for the control of vehicle 800, artificial intelligence of vehicle 800, infotainment and / or the like for vehicle 800.

[0097] Vehicle 800 may include one or more System-on-Chip (SoC) 804s. SoC 804 may include a CPU 806, GPU 808, processor 810, cache 812, accelerator 814, data storage 816, and / or other components and features not shown. SoC 804 can be used to control vehicle 800 across a wide variety of platforms and systems. For example, one or more SoCs 804s may be combined with an HD map 822 in a system (e.g., the system of vehicle 800), the HD map being transmitted via a network interface 824 from one or more servers (e.g., [server name missing]). Figure 8D One or more servers (878) receive map refresh and / or updates.

[0098] The CPU 806 may include CPU clusters or CPU complexes (or, alternatively, referred to herein as "CCPLEX"). The CPU 806 may include multiple cores and / or L2 cache. For example, in some embodiments, the CPU 806 may include eight cores in a coherent multiprocessor configuration. In some embodiments, the CPU 806 may include four dual-core clusters, each with a dedicated L2 cache (e.g., 2MB L2 cache). The CPU 806 (e.g., CCPLEX) may be configured to support simultaneous cluster operation, such that any combination of clusters of the CPU 806 can be active at any given time.

[0099] The CPU 806 can implement power management capabilities including one or more of the following features: automatic clock gating of hardware blocks when idle to conserve dynamic power; clock gating of each core when the core is not actively executing instructions due to the execution of WFI / WFE instructions; independent power gating of each core; independent clock gating of each core cluster when all cores are clock-gated or power-gated; and / or independent power gating of each core cluster when all cores are power-gated. The CPU 806 can further implement enhanced algorithms for managing power states, where allowed power states and desired wake-up times are specified, and the hardware / microcode determines the optimal power state to enter for the core, cluster, and CCPLEX. The processing core can support simplified power state entry sequences in software, with this work offloaded to the microcode.

[0100] GPU 808 can include an integrated GPU (alternatively referred to herein as an “iGPU”). GPU 808 can be programmable and efficient for parallel workloads. In some examples, GPU 808 can use an enhanced tensor instruction set. GPU 808 can include one or more streaming microprocessors, where each streaming microprocessor can include an LI cache (e.g., an LI 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 808 can include at least eight streaming microprocessors. GPU 808 can use a compute application programming interface (API). Additionally, GPU 808 can use one or more parallel computing platforms and / or programming models (e.g., NVIDIA’s CUDA).

[0101] In the case of automotive and embedded uses, GPU 808 can be power-optimized for best performance. For example, GPU 808 can be fabricated on a fin field-effect transistor (FinFET). However, this is not intended to be limiting, and GPU 808 can be fabricated using other semiconductor fabrication 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. Additionally, the streaming microprocessor can include independent parallel integer and floating point data paths to provide efficient execution of workloads with a mix of compute and address compute. The streaming microprocessor can include independent thread scheduling capabilities to allow for more fine-grained 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.

[0102] GPU 808 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, a synchronous graphics random access memory (SGRAM) can be used, such as a fifth generation graphics double data rate synchronous random access memory (GDDR5).

[0103] GPU 808 can include a unified memory technology that includes access counters 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 GPU 808 to access CPU 806 page tables directly. In such examples, when a GPU 808 memory management unit (MMU) experiences a miss, an address translation request can be transmitted to CPU 806. In response, CPU 806 can look up a virtual-to-physical mapping for the address in its page tables and transmit the translation back to GPU 808. In this way, the unified memory technology can allow a single unified virtual address space for memory of both CPU 806 and GPU 808, simplifying GPU 808 programming and porting applications to GPU 808.

[0104] In addition, GPU 808 can include access counters that can track how frequently GPU 808 accesses other processors’ memory. The access counters can help ensure that memory pages are migrated to the physical memory of the processor that accesses these pages most frequently.

[0105] SoC 804 can include any number of caches 812, including those described herein. For example, caches 812 can include an L3 cache available to both CPU 806 and GPU 808 (e.g., connected to both CPU 806 and GPU 808). Caches 812 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 4 MB or more, although smaller cache sizes can also be used.

[0106] SoC 804 can include an arithmetic logic unit (ALU) that can be utilized when performing processing (e.g., processing a DNN) in any of a variety of tasks or operations relative to vehicle 800. In addition, SoC 804 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 806 and / or GPU 808.

[0107] The SoC 804 can include one or more accelerators 814 (e.g., hardware accelerators, software accelerators, or a combination thereof). For example, the SoC 804 can include a hardware acceleration cluster that can include optimized hardware accelerators and / or large on-chip memory. This large on-chip memory (e.g., 4 MB SRAM) can enable the hardware acceleration cluster to accelerate neural networks and other computations. The hardware acceleration cluster can be used to supplement the GPU 808 and offload some of the tasks of the GPU 808 (e.g., freeing up more cycles of the GPU 808 for performing other tasks). As one example, the accelerators 814 can be used for targeted workloads (e.g., perception, convolutional neural networks (CNNs), etc.) that are stable enough to easily control acceleration. As used herein, the term “CNN” can include all types of CNNs, including region-based or region with convolutional neural networks (RCNNs) and fast RCNNs (e.g., for object detection).

[0108] The accelerators 814 (e.g., hardware acceleration cluster) 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 trillion operations per second for deep learning applications and inferencing. The TPUs can be accelerators that are configured to perform and optimized for performing image processing functions (e.g., for CNNs, RCNNs, etc.). 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 TPUs 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.

[0109] 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 limited to: CNNs for object recognition and detection using data from camera sensors; CNNs for distance estimation using data from camera sensors; CNNs for emergency vehicle detection and identification and detection using data from microphones; CNNs for face recognition and vehicle owner identification using data from camera sensors; and / or CNNs for safety and / or safety related events.

[0110] The DLA can perform any of the functions of the GPU 808, and by using an inferencing accelerator, the designer can target the DLA or the GPU 808 for any function. For example, the designer can focus the processing and floating point operations of CNNs on the DLA and leave other functions to the GPU 808 and / or other accelerators 814.

[0111] Accelerator 814 (e.g., hardware acceleration cluster) 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, for example and without limitation, any number of reduced instruction set computer (RISC) cores, direct memory access (DMA), and / or any number of vector processors.

[0112] 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 the 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 an instruction cache and / or a tightly coupled RAM.

[0113] The DMA can enable components of the PVA to access system memory independently of the CPU 806. The DMA can support any number of features 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 stride, vertical block stride, and / or depth stride.

[0114] The vector processors can be programmable processors 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 rate.

[0115] 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 acceleration 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 correction code (ECC) memory to enhance overall system security.

[0116] The accelerator 814 (e.g., hardware acceleration cluster) can include an on-chip computer vision network and SRAM to provide high bandwidth, low latency SRAM for the accelerator 814. In some examples, the on-chip memory 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 an on-chip computer vision network that interconnects the PVA and the DLA to the memory, for example using an APB.

[0117] The on-chip computer vision network 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.

[0118] In some examples, the SoC 804 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 functionality, 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.

[0119] The accelerator 814 (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.

[0120] 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.

[0121] 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.

[0122] 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 rather than false positive detections. For example, the system can set a threshold for confidence, and only consider detections that exceed the threshold to be 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 866 outputs related to vehicle 800 orientation, distance, 3D position estimates of objects obtained from the neural network and / or other sensors (e.g., LIDAR sensor 864 or RADAR sensor 860), etc.

[0123] SoC 804 can include one or more data stores 816 (e.g., memory). Data stores 816 can be on-chip memory of SoC 804, which can store neural networks to be executed on the GPU and / or DLA. In some examples, for redundancy and safety, data stores 816 can be large enough in capacity to store multiple instances of a neural network. Data stores 816 can include L2 or L3 cache 812. References to data stores 816 can include references to memory associated with PVA, DLA, and / or other accelerators 814 as described herein.

[0124] SoC 804 can include one or more processors 810 (e.g., embedded processors). The processors 810 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 804 boot sequence and can provide run-time power management services. The boot power and management processor can provide clock and voltage programming, auxiliary system low power state transitions, SoC 804 thermal and temperature sensor management, and / or SoC 804 power state management. Each temperature sensor can be implemented as a ring oscillator whose output frequency is proportional to temperature, and the SoC 804 can use the ring oscillator to detect the temperature of the CPU 806, GPU 808, and / or accelerator 814. If it is determined that the temperature exceeds a threshold, the boot and power management processor can enter a temperature fault routine and place the SoC 804 in a lower power state and / or place the vehicle 800 in a driver safe park mode (e.g., safely park the vehicle 800).

[0125] The processors 810 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.

[0126] The processors 810 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.

[0127] The processors 810 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.

[0128] The processors 810 can also include a real-time camera engine, which can include a specialized processor subsystem for handling real-time camera management.

[0129] The processor 810 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.

[0130] The processor 810 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 870, surround camera 874, 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.

[0131] 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.

[0132] 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 808 does not need to continuously render new surfaces. Even when the GPU 808 is powered on and active, doing 3D rendering, the video image compositor can be used to offload the GPU 808 to improve performance and responsiveness.

[0133] The SoC 804 can further include a Mobile Industry Processor Interface (MIPI) camera serial interface for receiving video and input from cameras, a high-speed interface, and / or a video input block that can be used for camera and related pixel input functions. The SoC 804 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.

[0134] The SoC 804 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 804 can be used to process data from cameras (connected over Gigabit Multimedia Serial Link and Ethernet), sensors (e.g., LIDAR sensors 864, RADAR sensors 860, etc. that can be connected over Ethernet), data from the bus 802 (e.g., speed of the vehicle 800, steering wheel position, etc.), data from GNSS sensors 858 (connected over Ethernet or CAN bus). The SoC 804 can also include dedicated high-performance mass storage controllers, which can include their own DMA engines, and which can be used to free up the CPU 806 from routine data management tasks.

[0135] The SoC 804 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 804 can be faster, more reliable, and even more energy and space efficient than conventional systems. For example, the accelerator 814, when combined with the CPU 806, GPU 808, and data storage 816, can provide a fast and efficient platform for level 3-5 autonomous vehicles.

[0136] The technology thus provides capabilities and functionality that cannot be achieved by 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 a requirement for practical level 3-5 autonomous vehicles.

[0137] In contrast to conventional systems, by providing a CPU complex, a GPU complex, and a hardware acceleration cluster, 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 820) 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 further 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.

[0138] 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 808.

[0139] 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 800. 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 804 provides security against theft and / or carjacking.

[0140] In another example, a CNN for emergency vehicle detection and recognition can use data from the microphones 896 to detect and recognize emergency vehicle sirens. In contrast to conventional systems that detect sirens using a general classifier and manually extract features, the SoC 804 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 858. Thus, for example, when operating in Europe, the CNN will seek to detect European sirens, and when in the United States, the CNN will seek to 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 862.

[0141] The vehicle can include a CPU 818 (e.g., a discrete CPU or dCPU) that can be coupled to the SoC 804 via a high-speed interconnect (e.g., PCIe). The CPU 818 can include, for example, an X86 processor. The CPU 818 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 804, and / or monitoring the status and health of the controller 836 and / or infotainment SoC 830.

[0142] The vehicle 800 can include a GPU 820 (e.g., a discrete GPU or dGPU) that can be coupled to the SoC 804 via a high-speed interconnect (e.g., NVIDIA’s NVLINK). The GPU 820 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 on input (e.g., sensor data) from sensors of the vehicle 800.

[0143] The vehicle 800 can also include a network interface 824 that can include one or more wireless antennas 826 (e.g., one or more wireless antennas for different communication protocols, such as cellular antennas, Bluetooth antennas, etc.). The network interface 824 can be used to enable wireless connections through the Internet with the cloud (e.g., with the server 878 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 800 with information about vehicles that are approaching the vehicle 800 (e.g., vehicles in front of, to the side of, and / or behind the vehicle 800). This functionality can be part of a cooperative adaptive cruise control functionality of the vehicle 800.

[0144] The network interface 824 can include a SoC that provides modulation and demodulation functionality and enables the controller 836 to communicate over a wireless network. The network interface 824 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.

[0145] The vehicle 800 can also include a data storage 828, which can include off-chip (e.g., off-SoC 804) storage. The data storage 828 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.

[0146] The vehicle 800 can also include a GNSS sensor 858. The GNSS sensor 858 (e.g., GPS and / or assisted GPS sensor) is used to assist in mapping, perception, occupancy grid generation, and / or path planning functions. Any number of GNSS sensors 858 can be used, including, for example and without limitation, a GPS using a USB connector with an Ethernet-to-serial (RS-232) bridge.

[0147] The vehicle 800 can also include a RADAR sensor 860. The RADAR sensor 860 can be used by the vehicle 800 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 860 can use the CAN and / or bus 802 (e.g., to transmit data generated by the RADAR sensor 860) 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 860 can be suitable for front, rear, and side RADAR use. In some examples, a pulsed Doppler RADAR sensor is used.

[0148] The RADAR sensor 860 can include different configurations, such as long-range with a narrow field of view, short-range with a wide field of view, short-range side coverage, and so on. In some examples, 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) implemented through two or more independent scans. The RADAR sensor 860 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 800 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 800.

[0149] As one example, a mid-range RADAR system can include a range of up to 860 m (front) or 80 m (rear) and a field of view of up to 42 degrees (front) or 850 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 blind spots to the sides of the vehicle.

[0150] A short-range RADAR system can be used in an ADAS system for blind spot detection and / or lane change assist.

[0151] The vehicle 800 can also include ultrasonic sensors 862. The ultrasonic sensors 862, which can be placed on the front, rear, and / or sides of the vehicle 800, can be used for parking assist and / or to create and update an occupancy grid. A wide variety of ultrasonic sensors 862 can be used, and different ultrasonic sensors 862 can be used for different detection ranges (e.g., 2.5 m, 4 m). The ultrasonic sensors 862 can operate at an ASIL B functional safety level.

[0152] The vehicle 800 can include LIDAR sensors 864. The LIDAR sensors 864 can be used for object and pedestrian detection, emergency braking, collision avoidance, and / or other functions. The LIDAR sensors 864 can be at an ASIL B functional safety level. In some examples, the vehicle 800 can include multiple LIDAR sensors 864 (e.g., two, four, six, etc.) that can use Ethernet (e.g., to provide data to a Gigabit Ethernet switch).

[0153] In some examples, the LIDAR sensors 864 can be capable of providing a list of objects and their distances for a 360-degree field of view. A commercially available LIDAR sensor 864 can have, for example, an advertised range of approximately 800 m, a precision of 2 cm - 3 cm, and support for an 800 Mbps Ethernet connection. In some examples, one or more flush-mounted LIDAR sensors 864 can be used. In such examples, the LIDAR sensors 864 can be implemented as small devices that can be embedded into the front, rear, sides, and / or corners of the vehicle 800. In such examples, the LIDAR sensors 864 can provide a field of view of up to 820 degrees horizontal and 35 degrees vertical with a range of 200 m, even for low reflectivity objects. Front-mounted LIDAR sensors 864 can be configured for a horizontal field of view between 45 degrees and 135 degrees.

[0154] 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 800. 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 fans. The flash LIDAR device can use 5 nanosecond Class I (eye-safe) laser pulses per frame and can capture the reflected laser light in the form of 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 864 can be less susceptible to motion blur, vibration, and / or jostling.

[0155] The vehicle can also include an IMU sensor 866. In some examples, the IMU sensor 866 can be located at the center of the rear axle of the vehicle 800. The IMU sensor 866 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 866 can include an accelerometer and a gyroscope, while in nine-axis applications, the IMU sensor 866 can include an accelerometer, a gyroscope, and a magnetometer.

[0156] In some embodiments, the IMU sensor 866 can be implemented as a microelectromechanical systems (MEMS) inertial navigation system (INS) that combines a microelectromechanical systems (MEMS) inertial sensor, a high-sensitivity GPS receiver, and an advanced Kalman filter algorithm to provide estimates of position, velocity, and attitude. As such, in some examples, the IMU sensor 866 can enable the vehicle 800 to estimate heading without input from a magnetic sensor by directly observing the change in velocity from GPS to the IMU sensor 866 and correlating it. In some examples, the IMU sensor 866 and the GNSS sensor 858 can be combined into a single integrated unit.

[0157] The vehicle can include a microphone 896 placed in and / or around the vehicle 800. The microphone 896 can be used for emergency vehicle detection and identification, among other things.

[0158] The vehicle can also include any number of camera types, including stereo cameras 868, wide-view cameras 870, infrared cameras 872, surround-view cameras 874, long and / or mid-range cameras 898, and / or other camera types. These cameras can be used to capture image data around the entire periphery of the vehicle 800. The types of cameras used depend on the embodiment and requirements of the vehicle 800, and any combination of camera types can be used to provide the necessary coverage around the vehicle 800. 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 without 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 8A and Figure 8B are described in more detail.

[0159] The vehicle 800 can also include vibration sensors 842. The vibration sensors 842 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 842 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).

[0160] The vehicle 800 can include an ADAS system 838. In some examples, the ADAS system 838 can include a SoC. The ADAS system 838 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.

[0161] The ACC system can use RADAR sensors 860, LIDAR sensors 864, and / or cameras. The ACC system can include longitudinal ACC and / or lateral ACC. Longitudinal ACC monitors and controls the distance to the vehicle immediately ahead of the vehicle 800 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 800. Lateral ACC is related to other ADAS applications such as LCA and CWS.

[0162] CACC uses information from other vehicles, which can be received from other vehicles via a wireless link via the network interface 824 and / or wireless antenna 826 or indirectly through a network connection, such as through the Internet. 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, such as vehicles immediately ahead of and in the same lane as the vehicle 800, 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 800, CACC can be more reliable, and it has the potential to improve traffic flow and reduce road congestion.

[0163] 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 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component. FCW systems can provide warnings in the form of, for example, sound, visual warnings, vibrations, and / or quick brake pulses.

[0164] 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 860 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.

[0165] LDW systems provide visual, audible, and / or tactile warnings, such as steering wheel or seat vibrations, to alert the driver when the vehicle 800 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 a display, speaker, and / or vibrating component.

[0166] An LKA system is a variation of the LDW system. If the vehicle 800 begins to leave the lane, the LKA system provides a steering input or brake to correct the vehicle 800.

[0167] 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 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component.

[0168] 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 800 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 860 coupled to a dedicated processor, DSP, FPGA, and / or ASIC that is electrically coupled to driver feedback such as a display, speaker, and / or vibrating component.

[0169] 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 800, in the case of conflicting results, the vehicle 800 itself must decide whether to heed the results from the primary computer or the secondary computer (e.g., the first controller 836 or the second controller 836). For example, in some embodiments, the ADAS system 838 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 838 can be provided to a supervisory MCU. If the outputs from the primary computer and the secondary computer conflict, the supervisory MCU must determine how to reconcile the conflict to ensure safe operation.

[0170] 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.

[0171] 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 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 the neural network with associated memory. In preferred embodiments, the supervisory MCU can include and / or be included as a component of the SoC 804.

[0172] In other examples, the ADAS system 838 can include secondary computers that perform ADAS functions using traditional computer vision rules. In this way, the secondary computers 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.

[0173] In some examples, the output of the ADAS system 838 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 838 indicates a forward collision warning due to an object immediately ahead, the perception block can use this information in identifying the object. In other examples, the secondary computer can have its own neural network that is trained and thus reduces the risk of false positives as described herein.

[0174] The vehicle 800 can also include an infotainment SoC 830 (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 830 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, Wi-Fi, etc.), and / or information services (e.g., a navigation system, park assist, a radio data system, vehicle-related information such as fuel level, total distance covered, brake fluid level, oil level, doors open / closed, air filter information, etc.) to the vehicle 800. For example, the infotainment SoC 830 can include a radio, a disc player, a navigation system, a video player, USB and Bluetooth connectivity, an in-car computer, in-car entertainment, Wi-Fi, steering wheel audio controls, hands-free voice controls, a heads-up display (HUD), an HMI display 834, a telematics device, a control panel (e.g., for controlling and / or interacting with various components, features, and / or systems), and / or other components. The infotainment SoC 830 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 838, autonomous driving information such as planned vehicle maneuvers, trajectories, surrounding environment information (e.g., intersection information, vehicle information, road information, etc.), and / or other information.

[0175] The infotainment SoC 830 can include GPU functionality. The infotainment SoC 830 can communicate with other devices, systems, and / or components of the vehicle 800 over the bus 802 (e.g., a CAN bus, Ethernet, etc.). In some examples, the infotainment SoC 830 can be coupled to a supervisory MCU such that, in the event of a failure of the host controller 836 (e.g., a primary and / or backup computer of the vehicle 800), the GPU of the infotainment system can perform some autonomous driving functions. In such examples, the infotainment SoC 830 can place the vehicle 800 in a driver safe park mode as described herein.

[0176] Vehicle 800 may also include an instrument cluster 832 (e.g., a digital instrument panel, electronic instrument cluster, digital instrument panel, etc.). The instrument cluster 832 may include a controller and / or a supercomputer (e.g., a discrete controller or supercomputer). The instrument cluster 832 may include a set of instruments such as a speedometer, fuel level, oil pressure, tachometer, odometer, turn indicator, shift position indicator, seatbelt warning light, parking brake warning light, engine malfunction indicator, airbag (SRS) system information, lighting controls, safety system controls, navigation information, etc. In some examples, information may be displayed and / or shared between the infotainment SoC 830 and the instrument cluster 832. In other words, the instrument cluster 832 may be included as part of the infotainment SoC 830, or vice versa.

[0177] Figure 8D For cloud-based servers and according to some embodiments of this disclosure Figure 8A This is a system diagram illustrating communication between example autonomous vehicles 800. System 876 may include server 878, network 890, and vehicles including vehicle 800. Server 878 may include multiple GPUs 884(A)-884(H) (collectively referred to herein as GPU 884), PCIe switches 882(A)-882(H) (collectively referred to herein as PCIe switch 882), and / or CPUs 880(A)-880(B) (collectively referred to herein as CPU 880). GPUs 884, CPUs 880, and PCIe switches may be interconnected with high-speed interconnects and / or PCIe connections 886, such as, but not limited to, NVLink interface 888 developed by NVIDIA. In some examples, GPUs 884 are connected via NVLink and / or NVSwitch SoCs, and GPUs 884 and PCIe switches 882 are connected via PCIe interconnects. Although eight GPUs 884, two CPUs 880, and two PCIe switches are shown in the diagram, this is not intended to be limiting. Depending on the embodiment, each of the servers 878 may include any number of GPUs 884, CPUs 880, and / or PCIe switches. For example, each of the servers 878 may include eight, sixteen, thirty-two, and / or more GPUs 884.

[0178] The server 878 can receive image data from vehicles over the network 890 and representing images showing unexpected or changing road conditions such as a road work that recently started. The server 878 can transmit neural networks 892, updated neural networks 892, and / or map information 894, including information about traffic and road conditions, to vehicles over the network 890. Updates to the map information 894 can include updates to the HD map 822, e.g., information about construction sites, potholes, curves, floods, or other obstacles. In some examples, the neural networks 892, updated neural networks 892, and / or map information 894 can have been generated from experience using training performed at a data center (e.g., using the server 878 and / or other servers) and / or from data received from any number of vehicles in the environment.

[0179] The server 878 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 variations 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 890) and / or the machine learning models can be used by the server 878 to remotely monitor vehicles.

[0180] In some examples, the server 878 can receive data from vehicles and apply the data to the latest real-time neural networks for real-time intelligent inference. The server 878 can include deep learning supercomputers and / or specialized AI computers powered by GPUs 884, such as the DGX and DGX Station machines developed by NVIDIA. However, in some examples, the server 878 can include deep learning infrastructure of a data center that is powered by CPUs only.

[0181] The deep learning infrastructure of the server 878 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 800. For example, the deep learning infrastructure can receive periodic updates from the vehicle 800, such as a sequence of images and / or objects located in the sequence of images that the vehicle 800 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 800, and if the results do not match and the infrastructure concludes that the AI in the vehicle 800 is malfunctioning, then the server 878 can transmit a signal to the vehicle 800 instructing the fail-safe computer of the vehicle 800 to take control, notify the passengers, and complete a safe parking operation.

[0182] For inference, the server 878 can include GPUs 884 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 responses possible. In other examples, such as where performance is less important, CPU-, FPGA-, and other processor-powered servers can be used for inference.

[0183] Example Computing Device

[0184] Figure 9 A block diagram of one or more example computing devices 900 suitable for implementing some embodiments of the present disclosure is provided. The computing device 900 can include an interconnection system 902 coupling the following devices: a memory 904, one or more central processing units (CPU) 906, one or more graphics processing units (GPU) 908, a communication interface 910, an input / output (I / O) port 912, an input / output component 914, a power supply 916, one or more presentation components 918 (e.g., a display), and one or more logic units 920.

[0185] Although Figure 9 Various blocks of the computing device 900 are shown as connected via the interconnection system 902 with a bus, this is not intended to be limiting and is for clarity only. For example, in some embodiments, a presentation component 918 such as a display device can be considered an I / O component 914 (e.g., if the display is a touchscreen). As another example, a CPU 906 and / or GPU 908 can include memory (e.g., the memory 904 can represent a storage device in addition to the memory of the GPU 908, CPU 906, and / or other components). In other words, the memory 904 can be a component of the computing device 900 that is separate from the GPU 908, CPU 906, and / or other components, or the memory 904 can be integrated into the GPU 908, CPU 906, and / or other components. Figure 9The computing device 900 is intended to represent various computing devices used by the techniques described herein. For example, the computing device 900 can represent a desktop computer, a laptop computer, a server, a mobile device, a game console, an electronic control unit (ECU), a virtual reality system, and / or other device or system types. The techniques described herein are not limited to use by any particular type of computing device. Figure 9 The computing device 900 is intended to represent various computing devices used by the techniques described herein. For example, the computing device 900 can represent a desktop computer, a laptop computer, a server, a mobile device, a game console, an electronic control unit (ECU), a virtual reality system, and / or other device or system types. The techniques described herein are not limited to use by any particular type of computing device.

[0186] The interconnect system 902 can represent one or more links or buses, such as an address bus, a data bus, a control bus, or a combination thereof. The interconnect system 902 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 906 can be directly connected to the memory 904. Further, the CPU 906 can be directly connected to the GPU 908. Where there are direct connections or point-to-point connections between components, the interconnect system 902 can include a PCIe link to perform the connection. In these examples, a PCI bus need not be included in the computing device 900.

[0187] The memory 904 can include any of a wide variety of computer-readable media. The computer-readable media can be any available media that can be accessed by the computing device 900. The 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.

[0188] The computer storage media can include volatile and nonvolatile, 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, the memory 904 can store computer-readable instructions (e.g., which represent programs and / or program elements, such as an operating system). The 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 that can be used to store the desired information and that can be accessed by the computing device 900. The computer storage media

[0189] Computer storage media can include computer-readable instructions, data structures, program modules, and / or other data 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 above should also be included within the scope of computer readable media.

[0190] The CPUs 906 can be configured to execute at least some of the computer-readable instructions in order to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. Each of the CPUs 906 can include one or more cores (e.g., one, two, four, eight, twenty-eight, seventy-two, etc.) capable of handling a large number of software threads concurrently. The CPUs 906 can include any type of processors and can include different types of processors depending on the type of computing device 900 being implemented (e.g., processors with fewer cores for mobile devices and processors with more cores for servers). For example, depending on the type of computing device 900, the processors can be Advanced RISC Machines (ARM) processors implemented using reduced instruction set computing (RISC) or x86 processors implemented using complex instruction set computing (CISC). The computing device 900 can include one or more CPUs 906 in addition to one or more microprocessors or supplemental co-processors such as math co-processors.

[0191] In addition to or instead of CPU 906, GPU 908 can be configured to execute at least some computer-readable instructions to control one or more components of computing device 900 to perform one or more of the methods and / or processes described herein. One or more of GPUs 908 can be integrated GPUs (e.g., one or more of CPUs 906 and / or one or more of GPUs 908 can be discrete GPUs). In embodiments, one or more of GPUs 908 can be a co-processor of one or more of CPUs 906. GPUs 908 can be used by computing device 900 to render graphics (e.g., 3D graphics) or to perform general purpose computing. For example, one or more GPUs 908 can be used for general purpose computing on GPUs (GPGPU). GPUs 908 can include hundreds or thousands of cores capable of processing hundreds or thousands of software threads concurrently. GPUs 908 can generate pixel data for output images in response to rendering commands (e.g., rendering commands from CPUs 906 received via a host interface). GPUs 908 can include graphics memory, such as display memory, for storing pixel data or any other suitable data (e.g., GPGPU data). Display memory can be included as part of memory 904. GPUs 908 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 908 can generate different portions of pixel data or GPGPU data for output or for different outputs (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.

[0192] In addition to or instead of the CPU 906 and / or GPU 908, the logic unit 920 can be configured to execute at least some computer-readable instructions to control one or more components of the computing device 900 to perform one or more of the methods and / or processes described herein. In embodiments, the one or more CPUs 906, the one or more GPUs 908, and / or the one or more logic units 920 can perform any combination of the methods, processes, and / or portions thereof, discretely or jointly. One or more of the logic units 920 can be part of and / or integrated in one or more of the CPUs 906 and / or GPUs 908, and / or one or more of the logic units 920 can be discrete components or otherwise external to the CPUs 906 and / or GPUs 908. In embodiments, one or more of the logic units 920 can be a co-processor of one or more of the CPUs 906 and / or one or more of the GPUs 908.

[0193] Examples of the logic unit 920 include one or more processing cores and / or components thereof, such as a tenant core (TC), a tenant processing unit (TPU), a pixel visual core (PVC), a visual processing unit (VPU), a graphics processing cluster (GPC), a texture processing cluster (TPC), a streaming multiprocessor (SM), a tree traversal unit (TTU), an artificial intelligence accelerator (AIA), a deep learning accelerator (DLA), an arithmetic logic unit (ALU), an application-specific integrated circuit (ASIC), a floating-point unit (FPU), an input / output (I / O) element, a peripheral component interconnect (PCI) or peripheral component interconnect express (PCIe) element, and the like.

[0194] The communication interface 910 can include one or more receivers, transmitters, and / or transceivers that enable the computing device 900 to communicate with other computing devices via electronic communication networks, including wired and / or wireless communication. The communication interface 910 can include components and functionality to enable communication over any of a number of different networks, such as wireless networks (e.g., Wi-Fi, Z-Wave, Bluetooth, Bluetooth LE, ZigBee, and the like), wired networks (e.g., communication over Ethernet or InfiniBand), low-power wide-area networks (e.g., LoRaWAN, SigFox, and the like), and / or the Internet.

[0195] I / O ports 912 can enable the computing device 900 to logically couple to other devices including I / O components 914, presentation components 918, and / or other components, some of which can be built into (e.g., integrated with) the computing device 900. Illustrative I / O components 914 include a microphone, mouse, keyboard, joystick, game pad, game controller, satellite dish, scanner, printer, wireless device, etc. The I / O components 914 can provide a natural user interface (NUI) that processes air gestures, voice, or other physiological inputs generated by a user. In some instances, inputs can be transmitted to an appropriate network element for further processing. A NUI can implement any combination of speech recognition, stylus recognition, facial recognition, biometric recognition, gesture recognition both on screen and adjacent to the screen, air gestures, head and eye tracking, and touch recognition (as described in more detail below) associated with a display of the computing device 900. The computing device 900 can include depth cameras, infrared cameras, RGB cameras, touch screens, and combinations of these, such as stereoscopic camera systems to provide a natural user interface (NUI) for computer interaction. Additionally, the computing device 900 can include accelerometers or gyroscopes (e.g., as part of an inertial measurement unit (IMU)) to enable detection of motion. In some examples, the output of the accelerometers or gyroscopes can be used by the computing device 900 to render immersive augmented reality or virtual reality.

[0196] A power supply 916 can include a hard-wired power supply, a battery power supply, or a combination thereof. The power supply 916 can supply power to the computing device 900 to enable the components of the computing device 900 to operate.

[0197] The presentation components 918 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 918 can receive data from other components (e.g., the GPU 908, the CPU 906, etc.) and output the data (e.g., as a image, video, sound, etc.).

[0198] The present disclosure can be described in the general context of machine-executable instructions, including computer-executable instructions, that are executed by a computer or other machines such as a personal digital assistant or other handheld device. Generally, program modules including routines, programs, objects, components, data structures, etc., refer to code that performs particular tasks or implements particular abstract data types. The present disclosure can be practiced in a variety of system configurations, including hand-held devices, consumer electronics, general- purpose computers, more specialty computing devices, and the like. 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.

[0199] As used herein, the term "and / or," with respect to a listing of two or more elements, means that at least one of the listed elements is present. For example, "A and / or B" is intended to cover A only, B only, or both A and B. Further, "at least one of A or B" is intended to cover at least one of A, at least one of B, or at least one of both A and B.

[0200] 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 inventors have contemplated that the claimed subject matter might also be embodied in other ways, to include different steps or combinations of steps similar to the ones described in this document, in conjunction with other present or future technologies. Moreover, although the terms "step" and / or "block" might 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 when the order of individual steps is explicitly described.

Claims

1. A method comprising: receiving image data representing an image generated using a first image sensor having a first field of view; generating updated image data by virtually adjusting the first field of view of the first image sensor such that a virtual center of the first image sensor is aligned with a horizontal line, generating projection image data representing a projected image based on the updated image data, applying the projection image data to a neural network trained to detect features represented by training image data representing images generated using one or more second image sensors having a second field of view smaller than the first field of view, and computing data representing feature detections corresponding to one or more features using the neural network and based at least in part on the projection image data.

2. The method of claim 1, wherein the first image sensor corresponds to a fisheye camera of a vehicle having a field of view equal to or greater than 120 degrees.

3. The method of claim 1, wherein, The generating updated image data includes generating updated image data corresponding to the virtually adjusted field of view based at least in part on the image data, and applying the updated image data to a stereographic projection algorithm.

4. The method of claim 3, wherein the virtually adjusting includes adjusting the first field of view in a vertical direction.

5. The method of claim 1, wherein the method further comprises: converting the feature detections to respective image-space locations corresponding to the projected image; and determining corresponding world-space locations of the one or more features corresponding to the feature detections from the image-space locations.

6. The method of claim 1, wherein the features correspond to at least one of a road layout, an object, or a wait condition.

7. A method comprising: receiving image data representing an image generated using an image sensor having a field of view; virtually adjusting the field of view of the image sensor such that a virtual center of the image sensor is aligned with a horizontal line; generating updated image data corresponding to the virtually adjusted field of view based at least in part on the image data, applying the updated image data to a stereographic projection algorithm to generate projection image data representing a projected image, applying the projection image data to one of a machine learning model or a computer vision algorithm, and computing data representing feature detections corresponding to one or more features using the one of the machine learning model or the computer vision algorithm and based at least in part on the projection image data.

8. The method of claim 7, wherein the machine learning model or the computer vision algorithm is trained or programmed based on training image data representing images generated using one or more image sensors having a field of view smaller than the field of view.

9. The method of claim 7, further comprising: sending data representing feature detections to a component of a vehicle for use by the vehicle in navigating a driving surface.

10. The method of claim 7, wherein virtually adjusting the field of view of the image sensor includes vertically adjusting the field of view in an upward direction.

11. The method of claim 7, wherein virtually adjusting the field of view of the image sensor comprises rotating a virtual sphere associated with the field of view of the image sensor by a predetermined number of degrees.

12. The method of claim 7, wherein the stereographic projection algorithm projects pixels of the updated image data onto a two-dimensional (2D) target plane to generate the projected image.

13. The method of claim 12, wherein each projected pixel of the projected image data corresponds to a pixel of the updated image data based at least in part on an intersection of a virtual line connecting a nadir of a virtual sphere associated with the image sensor and the projected pixel on the 2D target plane and a pixel of the updated image data corresponding to a spherical pixel.

14. The method of claim 7, wherein the method further comprises: converting the feature detections to respective image-space locations corresponding to the projected image; and determining, from the image-space locations, corresponding world-space locations of the one or more features corresponding to the feature detections.

15. The method of claim 7, wherein a field of view of the image sensor is greater than or equal to 120 degrees.

16. A system comprising: an image sensor having a field of view greater than 120 degrees and oriented at a first angle relative to a horizontal plane; a computing device comprising one or more processing devices and one or more memory devices communicatively coupled to the one or more processing devices, the one or more memory devices storing programming instructions that, when executed by the processing devices, cause instantiation of: a field of view adjuster to virtually adjust the field of view of the image sensor such that a virtual center of the image sensor is aligned with a horizontal line and generate updated image data from image data generated using the image sensor based at least in part on the virtually adjusted field of view; a projector to apply the updated image data to a stereographic projection algorithm to generate projected image data representing a projected image; and a feature detector to apply the projected image data to a machine learning model and compute, using the machine learning model and based at least in part on the projected image data, data representing feature detections corresponding to one or more features represented in the projected image.

17. The system of claim 16, wherein the machine learning model is trained based on training image data representing images generated using one or more image sensors having a field of view smaller than the field of view.

18. The system of claim 16, wherein the system further comprises an output converter to convert the feature detections to respective image-space locations corresponding to the image represented by the projected image data and determine, from the image-space locations, corresponding world-space locations of the one or more features corresponding to the feature detections.

19. The system of claim 18, wherein a field of view of the image sensor is greater than 180 degrees.

Images