Bird-eye view semantic segmentation label generation method based on inverse perspective transformation and point cloud projection

Abstract

The invention relates to an aerial view semantic segmentation label generation method based on inverse perspective transformation and point cloud projection, and the method comprises the steps: data collection: synchronizing camera and laser radar data at the same moment through a synchronization signal, and enabling the time stamp difference of all camera and laser radar sensor data at each moment not to exceed a set value; data labeling: performing joint labeling on m images and n point cloud data at the same moment, marking a static area of the road surface on the images, and marking a dynamic object 3D bounding box on the point cloud; the method comprises the following steps: generating a road surface region of a BEV label through inverse perspective transformation, perspective-perspective-based on the inverse perspective transformation of affine geometry, perspective-perspective the semantic segmentation label of the road surface at each camera visual angle to a BEV canvas, splicing the semantic segmentation labels, and carrying out refined processing on a spliced picture. According to the method, the accurate aerial view semantic segmentation label is directly generated from the original image and the point cloud which are synchronized at a certain moment, so that the aerial view is prevented from being acquired and labeled in a way of aerial photography of a road surface by an unmanned aerial vehicle, and the cost is reduced.

Description

technical field [0001] The invention relates to the technical field of surround perception for automatic driving of automobiles, in particular to a method for generating semantic segmentation tags for bird's-eye view images based on inverse perspective transformation and point cloud projection. Background technique [0002] The automatic driving system is one of the core systems of the current smart car. It is mainly composed of three large modules, namely the perception fusion module, the decision planning module, and the control module. Among them, the perception fusion is the front module of the other two modules. The accuracy of perception will directly determine the performance of the entire autonomous driving system. [0003] At present, many autonomous driving companies are beginning to focus on surround perception, specifically, distributing multiple cameras around the body, such as figure 1 As shown, it is the most common arrangement. These six cameras collect imag...

AI Technical Summary

Problems solved by technology

[0007] The purpose of the present invention is to provide a method for generating semantic segmentation tags for bird's-eye view images based on inverse perspective transformation and point cloud projection. The technical problem to be solved: there are currently two ways to obtain BEV tags in the industry. Then, the corresponding BEV tags are generated through the semantic information elements of the HD map. There are two defects in this method. First, the BEV tags are easily affected by the accuracy of the HD map. Second, the acquisition cost of the HD map is high and the cycle is long. , and the geographical scope of high-precision maps is limited, which restricts data diversity

Images