A novel early warning strategy for right-turning blind zone based on vulnerable road users detection

Blind zone detection of vehicles, as an essential function of Advanced Driver Assistance System, can effectively reduce the occurrence of traffic accidents and has attracted unprecedented attention. This paper develops an active collision avoidance method for right-turning blind zone, based on vulnerable road users (VRUs) detection. The proposed strategy consists of three main steps. First of all, an improved YOLOv4-tiny algorithm based on deep learning, combining two optimization strategies, is proposed to detect VRUs in right-turning blind zone more accurately and robustly. Secondly, a distance measurement method via monocular camera is used for ranging the distance between the host vehicle and the detected VRUs. Finally, a simple and effective vehicle active speed control algorithm is presented, based on distance and vehicle speed information, to provide early warning to the driver. This method was tested in a large driving dataset and in various actual driving situations. Experimental results show that, compared with the lightweight state-of-the-art methods, the improved YOLOv4-tiny has the best detection accuracy for VRUs and can stabilize at a detection speed of 50FPS on 1920*1080 resolution video, and that the measured distance error remains within 4%. A simulation test also proves that the proposed active speed control algorithm can effectively deliver early warning to drivers and avoid traffic accidents.