A method and system for point cloud and image fusion labeling

Abstract

The present invention relates to a method and system for point cloud and picture fusion labeling. The present invention first obtains the laser radar offset and rotation from the external parameters of the laser radar, and obtains the camera offset and rotation from the external parameters of the camera. , obtain the K matrix from the internal parameters of the camera; secondly, obtain the point cloud annotation frame set S from the point cloud annotation data, and the three-dimensional coordinate set PSi of the eight vertices of any annotation frame Bi in S; then use the lidar offset Quantity and rotation, camera offset and rotation, K matrix, calculate the 8 pixel coordinates corresponding to the 8 three-dimensional coordinates in PSi in the picture M, and form a set LSi; finally, according to the requirements, combine LSi to draw labels on the picture M frame, and assign corresponding attribute values ​​to the label frame. The labeling method proposed by the present invention can label point clouds and pictures at the same time, saving labeling time costs, manpower and material resources costs, and improving labeling efficiency.

Description

technical field [0001] The invention relates to the technical field of automatic driving, in particular to a method and system for fusion and labeling of point clouds and pictures used in model construction, training and optimization in the process of automatic driving research. Background technique [0002] In the process of autonomous driving research, a large amount of data such as pictures, point clouds and their corresponding ground truth values ​​are needed to build, train and optimize the model. The traditional labeling method is to separate the point cloud from the image and label them separately. This labeling method performs an independent labeling of the point cloud and the same object corresponding to the image, and the labeling efficiency is low. And there is no guarantee that the point cloud exactly matches the annotation attributes of the same object corresponding to the image. In addition, traditional image annotation methods often use rectangular or polygo...

AI Technical Summary

Problems solved by technology

This annotation method performs an independent annotation on the same object corresponding to the point cloud and the picture, and the annotation efficiency is low.

And there is no guarantee that the point cloud exactly matches the annotation attributes of the same object corresponding to the image

In addition, traditional image annotation methods often use rectangular or polygonal annotation frames to annotate the target object, which cannot obtain and display the real 3D data of the target object, which has created certain obstacles to the construction, training, and optimization of some models in autonomous driving research.

Images