Zoox, Inc.

David Pfeiffer

Zeng Wang

James William Vaisey Philbin

Allan Zelener

Robert Evan Mahieu

Dragomir Dimitrov Anguelov

Subhasis Das

Nitesh Shroff

Janek Hudecek

Yuanyuan Chen

Arthur Daniel Costea

Carden Taylor Bagwell

Qiang Zhai

Techniques for determining a probability that an object that has been appropriately labeled as a non-impeding object are disclosed. Non-impeding objects may be objects in an environment that do not substantially affect the operation of a vehicle in the environment, such as steam, fog, smoke, etc. A machine-learned model may classify such objects as non-impeding objects. The designation may be checked by using the lidar data and measurements associated with the lidar points associated with the object to determine a probability that the object is a non-impeding object. The lidar data may be aggregated within the individual voxels of a voxelized representation of the object for use in the probability determination.

Object detection soft consistency checker

Techniques for densifying annotations associated with sensor data are discussed herein. For example, techniques may include projecting a subset of lidar points into an unannotated two-dimensional image. Based on the projecting, the annotations of lidar points projected into the image can be associated with the image data. A subset of pixels of pixels can be dilated, and the dilated pixels can be determined to intersect (e.g., overlap, touch, etc.). Based on the pixels intersecting (or being adjacent) and being associated with the same segment identifier, a contour can be rendered around the corresponding non-dilated pixels to identify an object in the image data.

Annotating image data based on projected lidar data

Techniques for identifying lidar points associated with static objects, and using such lidar points to annotate objects within two-dimensional images are discussed herein. In some examples, an object manager may receive accumulations of lidar data captured from lidar devices of a vehicle while traversing within a driving environment. In some examples, the object manager may receive a plurality of annotated images. Such annotations may identify static objects within the driving environment. In some instances, the object manager may project a lidar point into an annotated image and determine that the lidar point is associated with an annotated pixel. Based on the pixel being associated with the annotated object, the object manager may determine that the lidar point is associated with object. In some examples, the object manager may determine a subset of lidar points that are associated with the object.

Lidar debris detection based on annotated image data

Techniques for detecting and classifying objects using lidar data are discussed herein. In some cases, the system may be configured to utilize a predetermined number of prior frames of lidar data to assist with detecting and classifying objects. In some implementations, the system may utilize a subset of the data associated with the prior lidar frames together with the full set of data associated with a current frame to detect and classify the objects.

Method for predicting behavior of object e as an autonomous vehicles, determining afggregated data based on first and second lidars data and transfer function

A vehicle can include various sensors to detect objects in an environment. In some cases, the object may be within a planned path of travel of the vehicle. In these cases, leaving the planned path may be dangerous to the passengers so the vehicle may, based on dimensions of the object, dimensions of the vehicle, and semantic information of the object, determine operational parameters associate with passing the object while maintaining a position within the planned path, if possible.

In-path obstacle detection and avoidance system

Techniques for identifying curbs are discussed herein. For instance, a vehicle may generate sensor data using one or more sensors, where the sensor data represents points associated with a driving surface and a sidewalk. The vehicle may then quantize the points into distance bins that are located laterally along the driving direction of the vehicle in order to generate spatial lines. Next, the vehicle may determine separation points for the spatial lines, where the separation points are configured to separate the points associated with the driving surface from the points associated with the sidewalk. The vehicle may then generate, using the separation points, a curve that represents the curb between the driving surface and the sidewalk. This way, the vehicle may use the curve while navigating, such as to avoid the curb and/or stop at a location that is proximate to the curb.

Techniques for identifying curbs

A system may include one or more processors configured to receive a plurality of images representing an environment. The images may include image data generated by an image capture device. The processors may also be configured to transmit the image data to an image segmentation network configured to segment the images. The processors may also be configured to receive sensor data associated with the environment including sensor data generated by a sensor of a type different than an image capture device. The processors may be configured to associate the sensor data with segmented images to create a training dataset. The processors may be configured to transmit the training dataset to a machine learning network configured to run a sensor data segmentation model, and train the sensor data segmentation model using the training dataset, such that the sensor data segmentation model is configured to segment sensor data.

Sensor data segmentation

Techniques for detecting an object in an environment and determining a probability that the object is a cloud of particulate matter. The cloud of particulate matter may include steam (e.g., emitted from a man-hole cover, a dryer exhaust port, etc.), exhaust from a vehicle (e.g., car, truck, motorcycle, etc.), environmental gases (e.g., resulting from sublimation, fog, evaporation, etc.), a cloud of dust, water splashing, blowing leaves, or other types of particulate matter that may be located in the environment of the vehicle and may not impact driving behavior (e.g., an autonomous vehicle may safely pass through the particulate matter without impact to the platform). A vehicle computing system may determine the probability that the object is a cloud of particulate matter and may control the vehicle based on the probability.

Identification of particulate matter in sensor data

Techniques for controlling a vehicle based on height data and/or classification data being determined utilizing multi-channel image data are discussed herein. The vehicle can capture lidar data as it traverses an environment. The lidar data can be associated with a voxel space as three-dimensional data. Semantic information can be determined and associated with the lidar data and/or the three-dimensional voxel space. A multi-channel input image can be determined based on the three-dimensional voxel space and input into a machine learned (ML) model. The ML model can output data to determine height data and/or classification data associated with a ground surface of the environment. The height data and/or classification data can be utilized to determine a mesh associated with the ground surface. The mesh can be used to control the vehicle and/or determine additional objects proximate the vehicle.

Estimating ground height based on lidar data

A vehicle can include various sensors to detect objects in an environment. Sensor data can be captured by a perception system in a vehicle and represented in a voxel space. Operations may include analyzing the data from a top-down perspective. From this perspective, techniques can associate and generate masks that represent objects in the voxel space. Through manipulation of the regions of the masks, the sensor data and/or voxels associated with the masks can be clustered or otherwise grouped to segment data associated with the objects.

Data segmentation using masks

Techniques are discussed for segmenting sensor data captured by a sensor to remove a ground surface from the sensor data. A first technique includes capturing sensor data represented as multichannel image and segmenting the image according to image processing techniques. The ground surface can be removed from the sensor data, and a subset of the sensor data can be associated with a voxel space. A second technique includes capturing sensor data and unprojecting the sensor data to generate three dimensional data, which can be associated with a voxel space. Ground plane data associated with a location can be accessed or determined and voxel data that is within a threshold height of the ground plane data can be removed from the voxel space. Clustering techniques can determine objects represented in the data, and a vehicle can be controlled based on the objects.

Time of flight data segmentation

A multiresolution voxel space is discussed herein. Data can be represented in individual levels in the multiresolution voxel space. A first level can correspond to a first region of an environment and a second level can correspond to a second region of an environment, the second region corresponding to a subset of the first region. In some examples, the levels can comprise a same number of voxels, such that the first level covers a large, low resolution region, while the second level covers a smaller, higher resolution region, though more levels are contemplated. Data represented in the voxel spaces can be processed at higher resolution where available and at a lower resolution where a data density is lower and combined in an intelligent manner. Voxel spaces can be updated based on movement of the sensor providing the data.

Growing community of inventors

Foster City, CA, United States of America

David Pfeiffer