Weather Classification for Lidar based on Deep Learning

Lidar is the most important sensor for roadside perception in autonomous driving and the Connected Automated Vehicle Highway(CAVH). Generally, perception algorithms based on point cloud only detect dynamic and static traffic participants, which lacks an analysis of the impact of abnormal weather types on point cloud detection. In practical applications, the CAVH system needs to determine whether it works within its operating design range according to different weather types, and adjusts accordingly. The main work of this paper is as follows: firstly, a large amout of various weather conditions data is collected as the basis for in-depth analysis of point cloud under changing environmental conditions. Secondly, the performance of roadside Lidar perception algorithm in different weather types is analyzed. Different from the traditional way of signal processing, this paper introduces deep neural network and realizes the classification of different weather types. Finally, in view of the efficiency problem of the classification network, an optimized structure is designed to realize the accurate identification of different weather types. The recognition accuracy rate increased to 96.86%, and the FPS increased to 30.