A multi-sensor fusion and object tracking algorithm for self-driving vehicles

Abstract

Vehicles need to detect threats on the road, anticipate emerging dangerous driving situations and take proactive actions for collision avoidance. Therefore, the study on methods of target detection and recognition are of practical value to a self-driving system. However, single sensor has its weakness, such as poor weather adaptability with lidar and camera. In this article, we propose a novel spatial calibration method based on multi-sensor systems, and the approach utilizes rotation and translation of the coordinate system. The validity of the proposed spatial calibration method is tested through comparisons with the data calibrated. In addition, a multi-sensor fusion and object tracking algorithm based on target level to detect and recognize targets is tested. Sensors contain lidar, radar and camera. The multi-sensor fusion and object tracking algorithm takes advantages of various sensors such as target location from lidar, target velocity from radar and target type from camera. Besides, multi-sensor fusion and object tracking algorithm can achieve information redundancy and increase environmental adaptability. Compared with the results of single sensor, this new approach is verified to have the accuracy of location, velocity and recognition by real data.