Bird-eye view semantic segmentation label generation method based on multi-frame semantic point cloud splicing

Abstract

The invention discloses an aerial view semantic segmentation label generation method based on multi-frame semantic point cloud splicing. The method comprises the following steps: 1) configuring six cameras and a laser radar on a vehicle; 2) calibrating an internal parameter of each camera and an external parameter relative to a vehicle body by using a calibration board, and calibrating an external parameter relative to the vehicle body of the laser radar; 3) synchronizing data acquired by the camera and the laser radar at the same moment; 4) performing joint labeling on six original images acquired by the camera at the same moment and the point cloud image of the laser radar; 5) the marked point cloud image is converted to each camera plane, and the semantic information of the image is utilized to dye the point cloud; and 6) splicing the continuous multi-frame semantic point clouds to a unified vehicle body coordinate system taking a certain frame as a benchmark, and projecting the continuous multi-frame semantic point clouds to a BEV canvas. The image semantic information and the point cloud information are utilized to generate the semantic point clouds for splicing, and finally the semantic point clouds are projected to the aerial view canvas for automatic generation, so that the cost of data labels is reduced.

Description

technical field [0001] The invention relates to the technical field of surround perception for automobile automatic driving, and in particular to a method for generating semantic segmentation tags for bird's-eye view images based on multi-frame semantic point cloud splicing. Background technique [0002] As a key technology of the new generation of smart cars, autonomous driving has been valued by more and more manufacturers. Generally speaking, the entire automatic driving system consists of three modules: perception fusion module, decision-making planning module, and control module. Among them, perception fusion is the front module of the other two modules, and its perception accuracy will directly determine the performance of the entire automatic driving system. performance. [0003] The current perception module technology is not limited to the traditional single forward camera (Forward Camera) configuration. Major manufacturers have begun to use multiple cameras around...

AI Technical Summary

Problems solved by technology

The biggest shortcoming of this method is that UAVs are usually subject to geographical control, so the scenes that can collect data will be limited

Furthermore, this acquisition method cannot be triggered through the shadow mode, making it difficult for the model to perform subsequent closed-loop iterations

Images