DE-102021113545-B4 - Method and system for operating a vehicle using a spatially invariant 3D convolution mesh via spherical coordinate input

DE102021113545B4DE 102021113545 B4DE102021113545 B4DE 102021113545B4DE-102021113545-B4

Abstract

Method for operating a vehicle (10), wherein the method comprises: Receiving a detection set (303, 305, 307) with respect to an object (50) at a radial distance from the vehicle (10); Selecting a convolution path (310, 320, 330) for the detection set (303, 305, 307) based on a radial distance of the object (50), wherein the convolution path (310, 320, 330) contains one or more convolution layers (312, 326, 336) and the number of one or more convolution layers depends on the radial distance of the object (50); Applying one or more convolution layers (312, 326, 336) of the selected convolution path (310, 320, 330) to the detection set (303, 305, 307) to generate a filtered data set; and Operating the vehicle (10) in relation to the object (50) using the filtered data set; characterized by the fact that the one or more convolution layers (312, 326, 336) contain a first convolution layer (312) and a sampling rate of the first convolution layer (312) that increases as the radial distance decreases; or the one or more convolution layers (312, 326, 336) contain a first convolution layer (312), the method further comprising defining a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) that is smaller than the first partial radial distance (302), applying the first convolution layer (312) with a first sampling rate when the object (50) is at the first partial radial distance (302), and applying the first convolution layer (312) with a second sampling rate when the object (50) is located at the second partial radial distance (304), where the second sampling rate is twice the first sampling rate; or the one or more convolution layers (312, 326, 336) contain a first convolution layer (312), the method further comprising defining a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) which is smaller than the first partial radial distance (302), and applying the first convolution layer (312) once to the detection set (303, 305, 307) in the first partial radial distance (302) or applying the first convolution layer (312) twice in the second partial radial distance (304), wherein a sampling rate of the first convolution layer (312) within the first partial radial distance (302) and the second partial radial distance (304) is the same; or The method further comprises: defining a first partial radial distance (302) and a second partial radial distance (304), defining a first weight over the first partial radial distance (302) and a second weight over the second partial radial distance (304) for selecting the folding path (310, 320, 330), wherein the first weight and the second weight taper off linearly with the radial distance when there is a radial boundary between the first partial radial distance (302) and the second partial radial distance (304), wherein the sum of the first weight and the second weight is equal to one.

Inventors

Oded Bialer
Amnon Jonas

Assignees

GM Global Technology Operations LLC

Dates

Publication Date: 20260513
Application Date: 20210526
Priority Date: 20201030

Claims (2)

Method for operating a vehicle (10), wherein the method comprises: obtaining a detection set (303, 305, 307) with respect to an object (50) at a radial distance from the vehicle (10); selecting a folding path (310, 320, 330) for the detection set (303, 305, 307) based on a radial distance of the object (50), wherein the folding path (310, 320, 330) includes one or more folding layers (312, 326, 336) and the number of the one or more folding layers depends on the radial distance of the object (50); Applying one or more convolution layers (312, 326, 336) of the selected convolution path (310, 320, 330) to the detection set (303, 305, 307) to generate a filtered data set; and operating the vehicle (10) with respect to the object (50) using the filtered data set; characterized in that the one or more convolution layers (312, 326, 336) include a first convolution layer (312) and a sampling rate of the first convolution layer (312) that increases as the radial distance decreases; or the one or more convolution layers (312, 326, 336) include a first convolution layer (312), the method further comprising defining a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) which is smaller than the first partial radial distance (302), applying the first convolution layer (312) with a first sampling rate when the object (50) is at the first partial radial distance (302), and applying the first convolution layer (312) with a second sampling rate when the object (50) is at the second partial radial distance (304), wherein the second sampling rate is twice the first sampling rate; or the one or more folding layers (312, 326, 336) contain a first folding layer (312), wherein the method further defines a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) which is smaller than the first partial radial distance (302), and includes applying the first convolution layer (312) once to the detection set (303, 305, 307) in the first partial radial distance (302) or applying the first convolution layer (312) twice in the second partial radial distance (304), wherein the sampling rate of the first convolution layer (312) is the same within the first partial radial distance (302) and the second partial radial distance (304); or the method further comprises: defining a first partial radial distance (302) and a second partial radial distance (304), defining a first weight over the first partial radial distance (302) and a second weight over the second partial radial distance (304) for selecting the folding path (310, 320, 330), wherein the first weight and the second weight taper off linearly with the radial distance when there is a radial boundary between the first partial radial distance (302) and the second partial radial distance (304), wherein the sum of the first weight and the second weight is equal to one.
System for operating a vehicle (10), the system comprising: a sensor for obtaining a detection set (303, 305, 307) with respect to an object (50) at a radial distance (301) from the vehicle (10); and a processor (36) configured to: select a convolution path (310, 320, 330) for the detection set (303, 305, 307) based on a radial distance (301) of the object (50), wherein the convolution path (310, 320, 330) contains one or more convolution layers (312, 326, 336) and the number of the one or more convolution layers (312, 326, 336) depends on the radial distance (301) of the object (50); Applying one or more convolution layers (312, 326, 336) of the selected convolution path (310, 320, 330) to the detection set (303, 305, 307) to generate a filtered data set; and operating the vehicle (10) with respect to the object (50) using the filtered data set, characterized in that the one or more convolution layers (312, 326, 336) include a first convolution layer (312) and a sampling rate of the first convolution layer (312) that increases as the radial distance (301) decreases; or the one or more convolutional layers (312, 326, 336) contain a first convolutional layer (312), wherein the processor (36) is further configured to define a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) which is smaller than the first partial radial distance (302), to apply the first convolutional layer (312) at a first sampling rate when the object (50) is at the first partial radial distance (302), and to apply the first convolutional layer (312) at a second sampling rate when the object (50) is at the second partial radial distance (304), wherein the second sampling rate is twice the first sampling rate; or the one or more convolutional layers (312, 326, 336) contain a first convolutional layer (312), and wherein the processor (36) is further configured to define a first partial radial distance (302) with respect to the vehicle (10) and a second partial radial distance (304) with respect to the vehicle (10) which is smaller than the first partial radial distance (302), and to perform the one-time application of the first convolutional layer (312) to the detection set (303, 305, 307) at the first partial radial distance (302) or the two-time application of the first convolutional layer (312) at the second partial radial distance (304), wherein a sampling rate of the first convolutional layer (312) within the first partial radial distance (302) and the second partial radial distance (304) is the same; or the processor (36) is further configured to define a first partial radial distance (302) and a second partial radial distance (304) and to define a first weight over the first partial radial distance (302) and a second weight over the second partial radial distance (304) for selecting the folding path (310, 320, 330), wherein the first weight and the second weight taper off linearly with the radial distance at a radial boundary between the first partial radial distance (302) and the second partial radial distance (304), wherein the sum of the first weight and the second weight is equal to one.

Description

INTRODUCTION The present invention relates to a system and a method for determining a feature of an object with respect to a vehicle, and in particular a system and a method according to the preamble of claim 1 or claim 1 for operating a vehicle, as essentially described in the article by ENGELS, G. [et al.] with the title "3D Object Detection from LiDAR Data using Distance Dependent Feature Extraction" (Proceedings of the 6th International Conference on Vehicle Technology and Intelligent Transport Systems, 2020, pp. 289-300. SCITEPRESS - Science and Technology Publications [online]. DOI: 10.5220/0009330402890300, In: SCITEPRESS ). An autonomous vehicle detects an object in its environment and navigates along a selected trajectory relative to the object. One or more sensors connected to the vehicle can be used to detect the object. Three-dimensional sensors, such as lidar and radar, essentially provide point cloud detections that can be used for convolutional meshes to extract features of the object. Because point cloud detections are received in a spherical coordinate system, the number of detections received from the object can vary significantly with distance. Since the convolutional mesh relies on spatially invariant features, applying standard convolutional meshes to such detections is problematic. Additionally, transforming the detections into Cartesian coordinates can produce a point density that varies with radial distance. Accordingly, it is desirable to create a convolutional mesh that provides an output that is spatially invariant with radial distance. SUMMARY According to an exemplary embodiment, a method for operating a vehicle is presented, characterized by the features of claim 1. A detection set is obtained with respect to an object at a radial distance from the vehicle. Based on the radial distance of the object, a convolution path for the detection set is selected, wherein the convolution path contains one or more convolution layers, and wherein the number of the one or more convolution layers depends on the radial distance of the object. The one or more convolution layers of the selected convolution path are applied to the detection set to generate a filtered data set. The vehicle is operated with respect to the object using the filtered data set. Furthermore, the method includes applying a second convolution layer at the first sampling rate to an intermediate set generated by the first convolution layer when the object is at the second partial radial distance. The method also includes defining a third partial radial distance with respect to the vehicle, which is smaller than the second partial radial distance; applying the first convolution layer at a third sampling rate when the object is at the third partial radial distance, where the third sampling rate is twice the second sampling rate; applying a second convolution layer at the second sampling rate to an intermediate set generated by the first convolution layer; and applying a third convolution layer to an output of the second convolution layer at the first sampling rate. Furthermore, according to the invention, a system for operating a vehicle is presented which is characterized by the features of claim 1. Furthermore, the processor is configured to apply a second convolution layer at the first sampling rate to an intermediate set generated by the first convolution layer when the object is at the second partial radial distance. The processor is also configured to define a third partial radial distance relative to the vehicle that is smaller than the second partial radial distance, to apply the first convolution layer at a third sampling rate when the object is at the third partial radial distance (where the third sampling rate is twice the second sampling rate), to apply a second convolution layer at the second sampling rate to an intermediate set generated by the first convolution layer, and to apply a third convolution layer to an output of the second convolution layer at the first sampling rate. Furthermore, a vehicle is described. The vehicle contains a sensor and a processor. The sensor receives a detection set with respect to an object at a radial distance from the vehicle. The processor is configured to select a convolution path for the detection set based on the radial distance of the object, wherein the convolution path contains one or more convolution layers, and wherein the number of the one or more convolution layers depends on the radial distance of the object. The processor then applies the one or more convolution layers of the selected convolution path to the detection set to generate a filtered data set and operates the vehicle with respect to the object using the filtered data set. In addition to one or more of the features described herein, the one or more convolutional layers include a first convolutional layer and a first convolutional layer sampling rate that increases as the radial distance decreases. Accordin