Nvidia Corporation

Joachim Pehserl

David Nister

Tilman Wekel

Sangmin Oh

Nikolai Smolyanskiy

Ruchi Bhargava

Ryan Oldja

Ibrahim Eden

Ke Chen

Alexander A Popov

David Wehr

Shane Murray

Gang Pan

Alexey Kamenev

Tae Eun Choe

Dong Zhang

Samuel Rupp Ogden

Baris Evrim Demiroz

Neda Cvijetic

Michael Brian Cox

Amir Akbarzadeh

Jacob Meyer

Ollin Boer Bohan

Hans Jonas Nilsson

Jens Christian Bo Joergensen

Aidin Ehsanibenafati

Anton Mitrokhin

Richard Whitcomb

Jake Guza

Marco Scoffier

Grant Monroe

An annotation pipeline may be used to produce 2D and/or 3D ground truth data for deep neural networks, such as autonomous or semi-autonomous vehicle perception networks. Initially, sensor data may be captured with different types of sensors and synchronized to align frames of sensor data that represent a similar world state. The aligned frames may be sampled and packaged into a sequence of annotation scenes to be annotated. An annotation project may be decomposed into modular tasks and encoded into a labeling tool, which assigns tasks to labelers and arranges the order of inputs using a wizard that steps through the tasks. During the tasks, each type of sensor data in an annotation scene may be simultaneously presented, and information may be projected across sensor modalities to provide useful contextual information. After all annotation tasks have been completed, the resulting ground truth data may be exported in any suitable format.

Ground truth data generation for deep neural network perception in autonomous driving applications

In various examples, perception of landmark shapes may be used for localization in autonomous systems and applications. In some embodiments, a deep neural network (DNN) is used to generate (e.g., per-point) classifications of measured 3D points (e.g., classified LiDAR points), and a representation of the shape of one or more detected landmarks is regressed from the classifications. For each of one or more classes, the classification data may be thresholded to generate a binary mask and/or dilated to generate a densified representation, and the resulting (e.g., dilated, binary) mask may be clustered into connected components that are iteratively: fitted a shape (e.g., a polynomial or Bezier spline for lane lines, a circle for top-down representations of poles or traffic lights), weighted, and merged. As such, the resulting connected components and their fitted shapes may be used to represent detected landmarks and used for localization, navigation, and/or other uses.

Landmark perception for localization in autonomous systems and applications

A deep neural network(s) (DNN) may be used to perform panoptic segmentation by performing pixel-level class and instance segmentation of a scene using a single pass of the DNN. Generally, one or more images and/or other sensor data may be stitched together, stacked, and/or combined, and fed into a DNN that includes a common trunk and several heads that predict different outputs. The DNN may include a class confidence head that predicts a confidence map representing pixels that belong to particular classes, an instance regression head that predicts object instance data for detected objects, an instance clustering head that predicts a confidence map of pixels that belong to particular instances, and/or a depth head that predicts range values. These outputs may be decoded to identify bounding shapes, class labels, instance labels, and/or range values for detected objects, and used to enable safe path planning and control of an autonomous vehicle.

Deep neural network for segmentation of road scenes and animate object instances for autonomous driving applications

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.

Deep neural network for detecting obstacle instances using radar sensors in autonomous machine applications

In various examples, systems and methods of the present disclosure detect and/or track objects in an environment using projection images generated from LiDAR. For example, a machine learning model—such as a deep neural network (DNN)—may be used to compute a motion mask indicative of motion corresponding to points representing objects in an environment. Various input channels may be provided as input to the machine learning model to compute a motion mask. One or more comparison images may be generated based on comparing depth values projected from a current range image to a coordinate space of a previous range image to depth values of the previous range image. The machine learning model may use the one or more projection images, the one or more comparison images, and/or the one or more range images to compute a motion mask and/or a motion vector output representation.

Dynamic object detection using LiDAR data for autonomous machine systems and applications

In various examples, a hazard detection system plots hazard indicators from multiple detection sensors to grid cells of an occupancy grid corresponding to a driving environment. For example, as the ego-machine travels along a roadway, one or more sensors of the ego-machine may capture sensor data representing the driving environment. A system of the ego-machine may then analyze the sensor data to determine the existence and/or location of the one or more hazards within an occupancy grid—and thus within the environment. When a hazard is detected using a respective sensor, the system may plot an indicator of the hazard to one or more grid cells that correspond to the detected location of the hazard. Based, at least in part, on a fused or combined confidence of the hazard indicators for each grid cell, the system may predict whether the corresponding grid cell is occupied by a hazard.

Hazard detection using occupancy grids for autonomous systems and applications

In various examples, systems and methods are disclosed that perform sensor fusion using rule-based and learned processing methods to take advantage of the accuracy of learned approaches and the decomposition benefits of rule-based approaches for satisfying higher levels of safety requirements. For example, in-parallel and/or in-serial combinations of early rule-based sensor fusion, late rule-based sensor fusion, early learned sensor fusion, or late learned sensor fusion may be used to solve various safety goals associated with various required safety levels at a high level of accuracy and precision. In embodiments, learned sensor fusion may be used to make more conservative decisions than the rule-based sensor fusion (as determined using, e.g., severity (S), exposure (E), and controllability (C) (SEC) associated with a current safety goal), but the rule-based sensor fusion may be relied upon where the learned sensor fusion decision may be less conservative than the corresponding rule-based sensor fusion.

Combining rule-based and learned sensor fusion for autonomous systems and applications

In various examples, a hazard detection system fuses outputs from multiple sensors over time to determine a probability that a stationary object or hazard exists at a location. The system may then use sensor data to calculate a detection bounding shape for detected objects and, using the bounding shape, may generate a set of particles, each including a confidence value that an object exists at a corresponding location. The system may then capture additional sensor data by one or more sensors of the ego-machine that are different from those used to capture the first sensor data. To improve the accuracy of the confidences of the particles, the system may determine a correspondence between the first sensor data and the additional sensor data (e.g., depth sensor data), which may be used to filter out a portion of the particles and improve the depth predictions corresponding to the object.

Particle-based hazard detection for autonomous machine

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.

Top-down object detection from LiDAR point clouds

Segmentation of lidar range images

In various examples, systems and methods are described that generate scene flow in 3D space through simplifying the 3D LiDAR data to '2.5D' optical flow space (e.g., x, y, and depth flow). For example, LiDAR range images may be used to generate 2.5D representations of depth flow information between frames of LiDAR data, and two or more range images may be compared to generate depth flow information, and messages may be passed—e.g., using a belief propagation algorithm—to update pixel values in the 2.5D representation. The resulting images may then be used to generate 2.5D motion vectors, and the 2.5D motion vectors may be converted back to 3D space to generate a 3D scene flow representation of an environment around an autonomous machine.

Belief propagation for range image mapping in autonomous machine applications

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.

Object detection and classification using LiDAR range images for autonomous machine applications

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.

Growing community of inventors

Lynnwood, WA, United States of America

Joachim Pehserl