Multi-sensor information fusion algorithm for night driving vehicle detection

Abstract

The ability to effectively detect vehicles and obtain their comprehensive information through sensors to achieve early warning of a collision and intelligent long-range light exposure are the basic prerequisites for safe night driving. However, sensors have their own performance characteristics, and a single type of sensor cannot fully perceive the environment. Therefore, a night driving vehicle-detection algorithm based on information fusion is proposed. An effective vehicle target priming algorithm based on millimeter-wave radar is designed that is experimentally shown to filter out a large number of irrelevant targets. A YOLOv5 benchmark model is optimized as follows: (a) intelligent data resampling addresses category imbalance; (b) head shared convolution addresses scale imbalance; and (c) lossless mosaic data enhancement solves the inconsistency between target features and labels generated by the original mosaic. A target matching algorithm correlates the detection results of a single sensor, providing comprehensive and reliable information including relative position and speed, category, and far light shield angle, laying a solid foundation for night driving safety.