Abstract: In various examples, past location information corresponding to actors in an environment and map information may be applied to a deep neural network (DNN)—such as a recurrent neural network (RNN)—trained to compute information corresponding to future trajectories of the actors. The output of the DNN may include, for each future time slice the DNN is trained to predict, a confidence map representing a confidence for each pixel that an actor is present and a vector field representing locations of actors in confidence maps for prior time slices. The vector fields may thus be used to track an object through confidence maps for each future time slice to generate a predicted future trajectory for each actor. The predicted future trajectories, in addition to tracked past trajectories, may be used to generate full trajectories for the actors that may aid an ego-vehicle in navigating the environment.

Abstract: In various examples, a deep neural network (DNN) is trained to accurately predict, in deployment, distances to objects and obstacles using image data alone. The DNN may be trained with ground truth data that is generated and encoded using sensor data from any number of depth predicting sensors, such as, without limitation, RADAR sensors, LIDAR sensors, and/or SONAR sensors. Camera adaptation algorithms may be used in various embodiments to adapt the DNN for use with image data generated by cameras with varying parameters—such as varying fields of view. In some examples, a post-processing safety bounds operation may be executed on the predictions of the DNN to ensure that the predictions fall within a safety-permissible range.

Abstract: Embodiments of the present disclosure relate to behavior planning for autonomous vehicles. The technology described herein selects a preferred trajectory for an autonomous vehicle based on an evaluation of multiple hypothetical trajectories by different components within a planning system. The various components provide an optimization score for each trajectory according to the priorities of the component and scores from multiple components may form a final optimization score. This scoring system allows the competing priorities (e.g., comfort, minimal travel time, fuel economy) of different components to be considered together. In examples, the trajectory with the best combined score may be selected for implementation. As such, an iterative approach that evaluates various factors may be used to identify an optimal or preferred trajectory for an autonomous vehicle when navigating an environment.

Abstract: In various examples, object fence corresponding to objects detected by an ego-vehicle may be used to determine overlap of the object fences with lanes on a driving surface. A lane mask may be generated corresponding to the lanes on the driving surface, and the object fences may be compared to the lanes of the lane mask to determine the overlap. Where an object fence is located in more than one lane, a boundary scoring approach may be used to determine a ratio of overlap of the boundary fence, and thus the object, with each of the lanes. The overlap with one or more lanes for each object may be used to determine lane assignments for the objects, and the lane assignments may be used by the ego-vehicle to determine a path or trajectory along the driving surface.

Abstract: In various examples, object fence corresponding to objects detected by an ego-vehicle may be used to determine overlap of the object fences with lanes on a driving surface. A lane mask may be generated corresponding to the lanes on the driving surface, and the object fences may be compared to the lanes of the lane mask to determine the overlap. Where an object fence is located in more than one lane, a boundary scoring approach may be used to determine a ratio of overlap of the boundary fence, and thus the object, with each of the lanes. The overlap with one or more lanes for each object may be used to determine lane assignments for the objects, and the lane assignments may be used by the ego-vehicle to determine a path or trajectory along the driving surface.

Abstract: In various examples, a current claimed set of points representative of a volume in an environment occupied by a vehicle at a time may be determined. A vehicle-occupied trajectory and at least one object-occupied trajectory may be generated at the time. An intersection between the vehicle-occupied trajectory and an object-occupied trajectory may be determined based at least in part on comparing the vehicle-occupied trajectory to the object-occupied trajectory. Based on the intersection, the vehicle may then execute the first safety procedure or an alternative procedure that, when implemented by the vehicle when the object implements the second safety procedure, is determined to have a lesser likelihood of incurring a collision between the vehicle and the object than the first safety procedure.

Abstract: In various examples, locations of directional landmarks, such as vertical landmarks, may be identified using 3D reconstruction. A set of observations of directional landmarks (e.g., images captured from a moving vehicle) may be reduced to 1D lookups by rectifying the observations to align directional landmarks along a particular direction of the observations. Object detection may be applied, and corresponding 1D lookups may be generated to represent the presence of a detected vertical landmark in an image.

Abstract: In various examples, a three-dimensional (3D) intersection structure may be predicted using a deep neural network (DNN) based on processing two-dimensional (2D) input data. To train the DNN to accurately predict 3D intersection structures from 2D inputs, the DNN may be trained using a first loss function that compares 3D outputs of the DNN—after conversion to 2D space—to 2D ground truth data and a second loss function that analyzes the 3D predictions of the DNN in view of one or more geometric constraints—e.g., geometric knowledge of intersections may be used to penalize predictions of the DNN that do not align with known intersection and/or road structure geometries. As such, live perception of an autonomous or semi-autonomous vehicle may be used by the DNN to detect 3D locations of intersection structures from 2D inputs.

Abstract: In various examples, sensor data may be collected using one or more sensors of an ego-vehicle to generate a representation of an environment surrounding the ego-vehicle. The representation may include lanes of the roadway and object locations within the lanes. The representation of the environment may be provided as input to a longitudinal speed profile identifier, which may project a plurality of longitudinal speed profile candidates onto a target lane. Each of the plurality of longitudinal speed profiles candidates may be evaluated one or more times based on one or more sets of criteria. Using scores from the evaluation, a target gap and a particular longitudinal speed profile from the longitudinal speed profile candidates may be selected. Once the longitudinal speed profile for a target gap has been determined, the system may execute a lane change maneuver according to the longitudinal speed profile.

Abstract: Systems and methods for performing visual odometry more rapidly. Pairs of representations from sensor data (such as images from one or more cameras) are selected, and features common to both representations of the pair are identified. Portions of bundle adjustment matrices that correspond to the pair are updated using the common features. These updates are maintained in register memory until all portions of the matrices that correspond to the pair are updated. By selecting only common features of one particular pair of representations, updated matrix values may be kept in registers. Accordingly, matrix updates for each common feature may be collectively saved with a single write of the registers to other memory. In this manner, fewer write operations are performed from register memory to other memory, thus reducing the time required to update bundle adjustment matrices and thus speeding the bundle adjustment process.

Abstract: In various examples, a deep neural network(s) (e.g., a convolutional neural network) may be trained to detect moving and stationary obstacles from RADAR data of a three dimensional (3D) space. In some embodiments, ground truth training data for the neural network(s) may be generated from LIDAR data. More specifically, a scene may be observed with RADAR and LIDAR sensors to collect RADAR data and LIDAR data for a particular time slice. The RADAR data may be used for input training data, and the LIDAR data associated with the same or closest time slice as the RADAR data may be annotated with ground truth labels identifying objects to be detected. The LIDAR labels may be propagated to the RADAR data, and LIDAR labels containing less than some threshold number of RADAR detections may be omitted. The (remaining) LIDAR labels may be used to generate ground truth data.

Abstract: In various examples, a deep neural network(s) (e.g., a convolutional neural network) may be trained to detect moving and stationary obstacles from RADAR data of a three dimensional (3D) space, in both highway and urban scenarios. RADAR detections may be accumulated, ego-motion-compensated, orthographically projected, and fed into a neural network(s). The neural network(s) may include a common trunk with a feature extractor and several heads that predict different outputs such as a class confidence head that predicts a confidence map and an instance regression head that predicts object instance data for detected objects. The outputs may be decoded, filtered, and/or clustered to form bounding shapes identifying the location, size, and/or orientation of detected object instances. The detected object instances may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.

Abstract: A deep neural network(s) (DNN) may be used to detect objects from sensor data of a three dimensional (3D) environment. For example, a multi-view perception DNN may include multiple constituent DNNs or stages chained together that sequentially process different views of the 3D environment. An example DNN may include a first stage that performs class segmentation in a first view (e.g., perspective view) and a second stage that performs class segmentation and/or regresses instance geometry in a second view (e.g., top-down). The DNN outputs may be processed to generate 2D and/or 3D bounding boxes and class labels for detected objects in the 3D environment. As such, the techniques described herein may be used to detect and classify animate objects and/or parts of an environment, and these detections and classifications may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.

Abstract: In various examples, object fence corresponding to objects detected by an ego-vehicle may be used to determine overlap of the object fences with lanes on a driving surface. A lane mask may be generated corresponding to the lanes on the driving surface, and the object fences may be compared to the lanes of the lane mask to determine the overlap. Where an object fence is located in more than one lane, a boundary scoring approach may be used to determine a ratio of overlap of the boundary fence, and thus the object, with each of the lanes. The overlap with one or more lanes for each object may be used to determine lane assignments for the objects, and the lane assignments may be used by the ego-vehicle to determine a path or trajectory along the driving surface.

Abstract: In various examples, locations of directional landmarks, such as vertical landmarks, may be identified using 3D reconstruction. A set of observations of directional landmarks (e.g., images captured from a moving vehicle) may be reduced to 1D lookups by rectifying the observations to align directional landmarks along a particular direction of the observations. Object detection may be applied, and corresponding 1D lookups may be generated to represent the presence of a detected vertical landmark in an image.

Abstract: An end-to-end system for data generation, map creation using the generated data, and localization to the created map is disclosed. Mapstreams—or streams of sensor data, perception outputs from deep neural networks (DNNs), and/or relative trajectory data—corresponding to any number of drives by any number of vehicles may be generated and uploaded to the cloud. The mapstreams may be used to generate map data—and ultimately a fused high definition (HD) map—that represents data generated over a plurality of drives. When localizing to the fused HD map, individual localization results may be generated based on comparisons of real-time data from a sensor modality to map data corresponding to the same sensor modality. This process may be repeated for any number of sensor modalities and the results may be fused together to determine a final fused localization result.

Abstract: An end-to-end system for data generation, map creation using the generated data, and localization to the created map is disclosed. Mapstreams—or streams of sensor data, perception outputs from deep neural networks (DNNs), and/or relative trajectory data—corresponding to any number of drives by any number of vehicles may be generated and uploaded to the cloud. The mapstreams may be used to generate map data—and ultimately a fused high definition (HD) map—that represents data generated over a plurality of drives. When localizing to the fused HD map, individual localization results may be generated based on comparisons of real-time data from a sensor modality to map data corresponding to the same sensor modality. This process may be repeated for any number of sensor modalities and the results may be fused together to determine a final fused localization result.

Abstract: An end-to-end system for data generation, map creation using the generated data, and localization to the created map is disclosed. Mapstreams—or streams of sensor data, perception outputs from deep neural networks (DNNs), and/or relative trajectory data—corresponding to any number of drives by any number of vehicles may be generated and uploaded to the cloud. The mapstreams may be used to generate map data—and ultimately a fused high definition (HD) map—that represents data generated over a plurality of drives. When localizing to the fused HD map, individual localization results may be generated based on comparisons of real-time data from a sensor modality to map data corresponding to the same sensor modality. This process may be repeated for any number of sensor modalities and the results may be fused together to determine a final fused localization result.

Abstract: In various examples, a deep neural network (DNN) may be used to detect and classify animate objects and/or parts of an environment. The DNN may be trained using camera-to-LiDAR cross injection to generate reliable ground truth data for LiDAR range images. For example, annotations generated in the image domain may be propagated to the LiDAR domain to increase the accuracy of the ground truth data in the LiDAR domain—e.g., without requiring manual annotation in the LiDAR domain. Once trained, the DNN may output instance segmentation masks, class segmentation masks, and/or bounding shape proposals corresponding to two-dimensional (2D) LiDAR range images, and the outputs may be fused together to project the outputs into three-dimensional (3D) LiDAR point clouds. This 2D and/or 3D information output by the DNN may be provided to an autonomous vehicle drive stack to enable safe planning and control of the autonomous vehicle.

Abstract: In various examples, object fence corresponding to objects detected by an ego-vehicle may be used to determine overlap of the object fences with lanes on a driving surface. A lane mask may be generated corresponding to the lanes on the driving surface, and the object fences may be compared to the lanes of the lane mask to determine the overlap. Where an object fence is located in more than one lane, a boundary scoring approach may be used to determine a ratio of overlap of the boundary fence, and thus the object, with each of the lanes. The overlap with one or more lanes for each object may be used to determine lane assignments for the objects, and the lane assignments may be used by the ego-vehicle to determine a path or trajectory along the driving surface.