Semi-supervised target labeling method and system for three-dimensional point cloud data

Abstract

The invention provides a semi-supervised target labeling method for three-dimensional point cloud data. The semi-supervised target labeling method comprises the following steps: reading original three-dimensional point cloud data of a current frame; preprocessing the original three-dimensional point cloud data to obtain obstacle point cloud data only containing an obstacle target; performing unsupervised target detection on the obstacle point cloud data to obtain an obstacle target 3D bounding box; using a multi-target tracking algorithm to automatically predict an obstacle target 3D boundingbox of the current frame according to the labeling result of the previous frame, and fusing the obstacle target 3D bounding box with an obstacle target 3D bounding box of unsupervised target detection; and checking and adjusting the fused 3D border of the obstacle target to obtain a final 3D bounding box of the current frame, i.e., a final labeling result of the current frame. The invention further provides a semi-supervised target labeling system for the three-dimensional point cloud data. According to the method, the problem of three-dimensional point cloud data annotation in the prior art can be well solved, and the annotation cost is reduced to the maximum extent.

Description

technical field [0001] The present invention relates to the field of computer technology, in particular to a point cloud data processing technology, in particular to a semi-supervised target labeling method and system for three-dimensional point cloud data. Background technique [0002] In recent years, lidar has played an important role in unmanned vehicle sensor solutions, using it to generate point cloud data to complete the perception of three-dimensional targets. Recent research work has demonstrated that deep learning-based methods have great application prospects in solving the problem of 3D object perception in LiDAR. However, deep learning solutions require large volumes of annotated 3D point cloud data. [0003] Labeling 3D point cloud data is currently facing great challenges, which are mainly summarized in two aspects. The first aspect is the complex labeling operation. Labeling 3D bounding boxes in 3D point clouds is more difficult than labeling 2D bounding box...

AI Technical Summary

Problems solved by technology

However, deep learning solutions require large volumes of annotated 3D point cloud data

[0003] Labeling 3D point cloud data is currently facing great challenges, which are mainly summarized in two aspects. The first aspect is the complex labeling operation. Labeling 3D bounding boxes in 3D point clouds is more difficult than labeling 2D bounding boxes in 2D images. It is much more complicated, because the three-dimensional coordinates, length, width, height, and orientation of the 3D bounding box need to be considered at the same time; the second aspect is repeated labeling operations. The point cloud data collected by LiDAR is usually provided in the form of sequence frames. There are differences but there will be a high degree of data association. If each frame is marked from scratch, it will lead to a large number of repeated labeling operations.

[0004] The current labeling methods and devices for 3D point cloud data have not been able to solve the above two problems well

The invention patent "Method and device for labeling point cloud data" published on April 20, 2018 with the publication number CN107945198A uses the feature recognition and tracking results of sensor data collected by sensors other than laser radar to compare the results collected by laser radar The segmentation and tracking results of the point cloud data are corrected, thereby reducing the manual workload of point cloud data labeling. However, this labeling method and device rely on other auxiliary sensors different from lidar, which increases the burden on users. At the same time, the auxiliary sensor acquisition The deviation of the data processing results will directly affect the final labeling results

Images