Zoox, Inc.

Benjamin Isaac Zwiebel

Subhasis Das

Dragomir Dimitrov Anguelov

William Anthony Silva

Qian Song

Juhana Kangaspunta

Kai Yu

Po-Jen Lai

Cheng-Hsin Wuu

James William Vaisey Philbin

Shida Shen

Techniques for performing deconvolution operations on data structures representing condensed sensor data are disclosed herein. Autonomous vehicle sensors can capture data in an environment that may include one or more objects. The sensor data may be processed by a convolutional neural network to generate condensed sensor data. The condensed sensor data may be processed by one or more deconvolution layers using a machine-learned upsampling transformation to generate an output data structure for improved object detection, classification, and/or other processing operations.

Learned deconvolutional upsampling decoding layer

Techniques for detecting and tracking objects in an environment are discussed herein. For example, techniques can include detecting a center point of a block of pixels associated with an object. Unimodal (e.g., Gaussian) confidence values may be determined for a group of pixels associated with an object. Proposed detection box center points may be determined based on the Gaussian confidence values of the pixels and an output detection box may be determined using filtering and/or suppression techniques. Further, a machine-learned model can be trained by determining parameters of a center pixel of the detection box and a focal loss based on the unimodal confidence value which can then be backpropagated to the other pixels of the detection.

Center-based detection and tracking

Techniques for a perception system of a vehicle that can detect and track objects in an environment are described herein. The perception system may include a machine-learned model that includes one or more different portions, such as different components, subprocesses, or the like. In some instances, the techniques may include training the machine-learned model end-to-end such that outputs of a first portion of the machine-learned model are tailored for use as inputs to another portion of the machine-learned model. Additionally, or alternatively, the perception system described herein may utilize temporal data to track objects in the environment of the vehicle and associate tracking data with specific objects in the environment detected by the machine-learned model. That is, the architecture of the machine-learned model may include both a detection portion and a tracking portion in the same loop.

Vehicle perception system with temporal tracker

Tracking a current and/or previous position, velocity, acceleration, and/or heading of an object using sensor data may comprise determining whether to associate a current object detection generated from recently received (e.g., current) sensor data with a previous object detection generated from formerly received sensor data. In other words, a track may identify that an object detected in former sensor data is the same object detected in current sensor data. However, multiple types of sensor data may be used to detect objects and some objects may not be detected by different sensor types or may be detected differently, which may confound attempts to track an object. An ML model may be trained to receive outputs associated with different sensor types and/or a track associated with an object, and determine a data structure comprising a region of interest, object classification, and/or a pose associated with the object.

Top-down view object detection and tracking

Techniques are discussed for controlling a vehicle, such as an autonomous vehicle, based on occluded areas in an environment. An occluded area can represent areas where sensors of the vehicle are unable to sense portions of the environment due to obstruction by another object. An occlusion grid representing the occluded area can be stored as map data or can be dynamically generated. An occlusion grid can include occlusion fields, which represent discrete two- or three-dimensional areas of driveable environment. An occlusion field can indicate an occlusion state and an occupancy state, determined using LIDAR data and/or image data captured by the vehicle. An occupancy state of an occlusion field can be determined by ray casting LIDAR data or by projecting an occlusion field into segmented image data. The vehicle can be controlled to traverse the environment when a sufficient portion of the occlusion grid is visible and unoccupied.

Occlusion aware planning and control

End-to-end vehicle perception system training

Object velocity and/or yaw rate detection and tracking

Techniques are disclosed for a combined machine learned (ML) model that may generate a track confidence metric associated with a track and/or a classification of an object. Techniques may include obtaining a track. The track, which may include object detections from one or more sensor data types and/or pipelines, may be input into a machine-learning (ML) model. The model may output a track confidence metric and a classification. In some examples, if the track confidence metric does not satisfy a threshold, the ML model may cause the suppression of the output of the track to a planning component of an autonomous vehicle.

Growing community of inventors

Burlingame, CA, United States of America

Benjamin Isaac Zwiebel