Nvidia Corporation

Sangmin Oh

David Nister

Minwoo Park

Junghyun Kwon

Joachim Pehserl

Zhaoting Ye

Yilin Yang

Bala Siva Sashank Jujjavarapu

Tilman Wekel

Nikolai Smolyanskiy

Ruchi Bhargava

Pekka Janis

Tommi Koivisto

Ryan Oldja

Daniel Herrera Castro

Alexander A Popov

Mehmet K Kocamaz

Dongwoo Lee

Ibrahim Eden

Hae-Jong Seo

Berta Rodriguez Hervas

Tae Eun Choe

Shane Murray

Dong Zhang

Wenchao Zheng

Baris Evrim Demiroz

Gang Pan

Ke Chen

Alexey Kamenev

David Weikersdorfer

Samuel Rupp Ogden

Emilie Lucie Eloïse Wirbel

Gyeong Woo Cheon

Aman Jhunjhunwala

Arthur Henry Rajala

Qian Lin

Bor-Jeng Chen

Neda Cvijetic

Michael Brian Cox

Jagadeesh Sankaran

Wanli Jiang

Hang Dou

Eric Viscito

Hans Jonas Nilsson

Siyi Li

Yichun Shen

Ke Xu

Kexuan Zou

Neeraj Sajjan

Aidin Ehsanibenafati

Chethan Ningaraju

Hao Tsui

Yongqing Liang

Daniel Per Olof Svensson

Bala Siva Jujjavarapu

In various examples, techniques for multi-dimensional tracking of objects using two-dimensional (2D) sensor data are described. Systems and methods may use first image data to determine a first 2D detected location and a first three-dimensional (3D) detected location of an object. The systems and methods may then determine a 2D estimated location using the first 2D detected location and a 3D estimated location using the first 3D detected location. The systems and methods may use second image data to determine a second 2D detected location and a second 3D detected location of a detected object, and may then determine that the object corresponds to the detected object using the 2D estimated location, the 3D estimated location, the second 2D detected location, and the second 3D detected location. The systems and method then generate, modify, delete, or otherwise update an object track that includes 2D state information and 3D state information.

Joint 2D and 3D object tracking for 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

A neural network may be used to determine corner points of a skewed polygon (e.g., as displacement values to anchor box corner points) that accurately delineate a region in an image that defines a parking space. Further, the neural network may output confidence values predicting likelihoods that corner points of an anchor box correspond to an entrance to the parking spot. The confidence values may be used to select a subset of the corner points of the anchor box and/or skewed polygon in order to define the entrance to the parking spot. A minimum aggregate distance between corner points of a skewed polygon predicted using the CNN(s) and ground truth corner points of a parking spot may be used simplify a determination as to whether an anchor box should be used as a positive sample for training.

Object detection using skewed polygons suitable for parking space detection

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, 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, one or more output channels of a deep neural network (DNN) may be used to determine assignments of obstacles to paths. To increase the accuracy of the DNN, the input to the DNN may include an input image, one or more representations of path locations, and/or one or more representations of obstacle locations. The system may thus repurpose previously computed information—e.g., obstacle locations, path locations, etc.—from other operations of the system, and use them to generate more detailed inputs for the DNN to increase accuracy of the obstacle to path assignments. Once the output channels are computed using the DNN, computed bounding shapes for the objects may be compared to the outputs to determine the path assignments for each object. Additionally, a machine may perform control operations based at least on the path assignments.

Obstacle to path assignment and performance of control operations based on assignment 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, techniques for sensor-fusion based object detection and/or free-space detection using ultrasonic sensors are described. Systems may receive sensor data generated using one or more types of sensors of a machine. In some examples, the systems may then process at least a portion of the sensor data to generate input data, where the input data represents one or more locations of one or more objects within an environment. The systems may then input at least a portion of the sensor data and/or at least a portion of the input data into one or more neural networks that are trained to output one or more maps or other output representations associated with the environment. In some examples, the map(s) may include a height, an occupancy, and/or height/occupancy map generated, e.g., from a birds-eye-view perspective. The machine may use these outputs to perform one or more operations.

Sensor fusion using ultrasonic sensors for autonomous systems and applications

In various examples, live perception from sensors of a vehicle may be leveraged to generate object tracking paths for the vehicle to facilitate navigational controls in real-time or near real-time. For example, a deep neural network (DNN) may be trained to compute various outputs—such as feature descriptor maps including feature descriptor vectors corresponding to objects included in a sensor(s) field of view. The outputs may be decoded and/or otherwise post-processed to reconstruct object tracking and to determine proposed or potential paths for navigating the vehicle.

Single and across sensor object tracking using feature descriptor mapping in 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

Apparatuses, systems, and techniques to perform non-maximum suppression (NMS) in parallel to remove redundant bounding boxes. In at least one embodiment, two or more parallel circuits to perform two or more portions of a NMS algorithm in parallel to remove one or more redundant bounding boxes corresponding to one or more objects within one or more digital images.

Performing non-maximum suppression in parallel

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.

Distance to obstacle detection in autonomous machine applications

In various examples, a deep neural network (DNN) is trained—using image data alone—to accurately predict distances to objects, obstacles, and/or a detected free-space boundary. The DNN may be trained with ground truth data that is generated using sensor data representative of motion of an ego-vehicle and/or sensor data from any number of depth predicting sensors—such as, without limitation, RADAR sensors, LIDAR sensors, and/or SONAR sensors. The DNN may be trained using two or more loss functions each corresponding to a particular portion of the environment that depth is predicted for, such that—in deployment—more accurate depth estimates for objects, obstacles, and/or the detected free-space boundary are computed by the DNN.

Distance estimation to objects and free-space boundaries in 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

San Jose, CA, United States of America

Sangmin Oh