CN-117058480-B - Tag data generation method, system, equipment and medium

CN117058480BCN 117058480 BCN117058480 BCN 117058480BCN-117058480-B

Abstract

The invention discloses a generation method, a system, equipment and a medium of tag data, wherein the generation method comprises the steps of acquiring an original foreground target point and an original background target point in a radar point cloud; generating a dense foreground target point and a dense background target point which correspond to each other based on the foreground target point and the background target point, respectively transforming the dense foreground target point and the dense background target point into a corresponding original target frame and an original position to obtain dense point clouds, voxelating the dense point clouds, and generating tag data corresponding to the dense point clouds. The method and the device not only solve the problem of missing label data in the 3D occupied network prediction task, but also avoid resource consumption caused by using manual labeling. The tag data generation method provided by the scheme is suitable for all outdoor road scene data sets meeting the conditions, and can provide a large amount of usable data for the research of 3D space occupation prediction tasks in the automatic driving field.

Inventors

YI TAO
XIA QING
Xiong Yineng
DING WENBO

Assignees

赛可智能科技(上海)有限公司

Dates

Publication Date: 20260505
Application Date: 20230726

Claims (8)

1. A method of generating tag data, the method comprising: Acquiring an original foreground target point and an original background target point in Lei Dadian cloud, and extracting ground points from the radar point cloud by adopting a cloth filtering algorithm; generating a corresponding dense foreground target point and a dense background target point through multi-frame densification based on the foreground target point and the background target point, wherein the generating of the dense foreground target point comprises the steps of transforming points in a 3D truth frame of each frame in the foreground target point to the same coordinate system, carrying out multi-frame aggregation, RANSAC point cloud registration and speed estimation on the transformed foreground target point, and turning points on one side, which are denser in the aggregated point cloud, to the other side through axisymmetric operation based on the characteristic that a vehicle is symmetric left and right along a vertical center line of a vehicle body; transforming the dense foreground target point and the dense background target point into corresponding original target frames and original positions respectively to obtain dense point clouds; and voxelizing the dense point cloud, and generating label data corresponding to the dense point cloud.
2. The method of generating tag data according to claim 1, wherein the step of generating the corresponding dense foreground target point and dense background target point includes: acquiring the foreground target point and performing multi-frame densification treatment to generate a dense foreground target point; And obtaining Lei Dadian ground points and the background target points in the cloud and performing multi-frame densification processing to generate dense background target points.
3. The method of generating tag data according to claim 1, wherein after the step of voxelizing the dense point cloud, the generating method comprises: and (3) performing densification operation on voxels corresponding to the dense point cloud by using a poisson surface reconstruction algorithm to generate target voxels.
4. The method of generating tag data according to claim 1, wherein after the step of voxelizing the dense point cloud, the generating method comprises: Acquiring a connecting line of each voxel corresponding to the dense point cloud and a camera optical center under each camera view angle; based on the connection, a camera visibility mask is generated.
5. The tag data generation method according to claim 1, wherein after the step of generating tag data corresponding to the dense point cloud, the generation method includes: and generating a corresponding occupied or unoccupied mask based on the label data.
6. A system for generating tag data, the system comprising: the first acquisition module is used for acquiring an original foreground target point and an original background target point in the radar point cloud, and extracting ground points from the radar point cloud by adopting a cloth filtering algorithm; The first generation module is used for generating a corresponding dense foreground target point and a dense background target point through multi-frame densification based on the foreground target point and the background target point, wherein the multi-frame densification process generates the dense foreground target point by transforming points in 3D truth frames of each frame of the foreground target point to the same coordinate system, carries out multi-frame aggregation, RANSAC point cloud registration and speed estimation on the transformed foreground target point, and turns over the points on one side, which are relatively dense, of the aggregated point cloud to the other side through axisymmetric operation based on the characteristic that a vehicle is symmetrical left and right along a vertical central line of a vehicle body; the transformation module is used for transforming the dense foreground target point and the dense background target point into a corresponding original target frame and an original position respectively so as to obtain a dense point cloud; And the second generation module is used for voxelizing the dense point cloud and generating label data corresponding to the dense point cloud.
7. An electronic device comprising a memory, a processor and a computer program stored on the memory for execution on the processor, wherein the processor implements the method of generating tag data according to any one of claims 1 to 5 when executing the computer program.
8. A computer-readable storage medium, on which a computer program is stored, characterized in that the computer program, when being executed by a processor, implements the steps of the tag data generation method according to any one of claims 1 to 5.

Description

Tag data generation method, system, equipment and medium Technical Field The disclosure relates to the technical field of automatic driving, and in particular relates to a method, a system, equipment and a medium for generating tag data. Background In the field of automatic driving, the perception of the surrounding environment is extremely critical, knowing the occupancy state of each voxel in a 3D (three-dimensional) space is beneficial to the vehicle to recognize and avoid various obstacles in an unknown environment, however, 3D occupancy network prediction is a complex task requiring a large amount of 3D scene semantic information for training, and if only sparse point cloud data is used for supervision, the network is difficult to predict the occupancy situation with sufficient density. The goal of the 3D occupancy network prediction task is to estimate the state of each voxel in the 3D scene, including occupancy, and semantic information, by a series of sensor input camera T-frame historical frame pictures. The occupation condition is occupation or idle, and the semantic information is the category of the object to which the voxel belongs. Therefore, for training data, a voxel level label is required that contains occupancy information and semantic information. If the manual labeling is directly carried out, a great deal of manpower is required to be consumed, and the efficiency is low. Disclosure of Invention The technical problem to be solved by the method is to overcome the defects that a great deal of manpower is required to be consumed and the efficiency is low in the prior art by directly manually marking, and to provide a method, a system, equipment and a medium for generating tag data. The technical problems are solved by the following technical scheme: The present disclosure provides a method for generating tag data, the method comprising: Acquiring an original foreground target point and an original background target point in Lei Dadian cloud; generating a corresponding dense foreground target point and a dense background target point based on the foreground target point and the background target point; transforming the dense foreground target point and the dense background target point into corresponding original target frames and original positions respectively to obtain dense point clouds; and voxelizing the dense point cloud, and generating label data corresponding to the dense point cloud. Preferably, the step of generating the corresponding dense foreground target point and dense background target point includes: acquiring the foreground target point and performing multi-frame densification treatment to generate a dense foreground target point; and obtaining the background target point and performing multi-frame densification processing to generate a dense background target point. Preferably, the step of acquiring the foreground target point and performing multi-frame densification processing includes: transforming the points of each frame in the foreground target point to the same coordinate system and performing multi-frame aggregation, mirror symmetry and point cloud registration; And/or, the step of acquiring the ground point and the background target point in the radar point cloud and performing multi-frame densification processing comprises the following steps: and performing multi-frame aggregation and least square method calibration on the point of each frame in the background target point. Preferably, the step of generating the corresponding dense foreground target point and dense background target point includes: acquiring the foreground target point and performing multi-frame densification treatment to generate a dense foreground target point; And obtaining Lei Dadian ground points and the background target points in the cloud and performing multi-frame densification processing to generate dense background target points. Preferably, after the step of voxelizing the dense point cloud, the generating method includes: and (3) performing densification operation on voxels corresponding to the dense point cloud by using a poisson surface reconstruction algorithm to generate target voxels. Preferably, after the step of voxelizing the dense point cloud, the generating method includes: Acquiring a connecting line of each voxel corresponding to the dense point cloud and a camera optical center under each camera view angle; based on the connection, a camera visibility mask is generated. Preferably, after the step of generating the tag data corresponding to the dense point cloud, the generating method includes: and generating a corresponding occupied or unoccupied mask based on the label data. The present disclosure provides a generation system of tag data, the generation system comprising: the first acquisition module is used for acquiring an original foreground target point and an original background target point in the radar point cloud; the first generation module is used for genera