Rear-end Collision Prevention System From Front Vehicle Based On Binocular Stereo Vision

With the continuous development of new energy vehicles and autonomous driving technology,people also put forward higher requirements for reliable driver assistance system.Forward Collision Warning(FCW)System is an important part of Advanced Driver Assistance System(ADAS).The system generally detects the front vehicle through the radar system and judges the distance,azimuth and relative speed of the front vehicle,but lacks the behavior analysis of the front vehicle.To solve this problem,this paper proposes a FCW system which combines vehicle speed measurement and vehicle behavior analysis.The binocular stereo vision system is carried on the target vehicle,and the speed of the vehicle moving in the same direction in front of the target vehicle is measured according to the binocular stereo vision principle.At the same time,the improved X3D-M behavior recognition algorithm is used to judge the state of the rear signal light of the front vehicle in the same direction,and analyze the possible behavior of the front vehicle.In this paper,a specific scheme is designed for the detection of vehicle tail characteristics,and it is used as the vehicle tracking and speed measurement module of FCW system.At the same time,the behavior recognition algorithm is designed and is used in the vehicle behavior analysis module of FCW system.The main work of this paper includes:(1)According to the characteristics of vehicles traveling in the same direction,this paper makes a data set of vehicle tail characteristics.In this data set,several YOLOv5 target detection algorithms with different structures are compared,and the YOLOv5 m network with small number of parameters and high detection accuracy is selected as the vehicle tail characteristic detection model.(2)Combined with the national standard GB4789-2019 on the external lighting installation and light signal devices for vehicles and trailers,the vehicle taillight status data set is made.Based on the data set,six behavior recognition algorithms based on deep learning are compared,and X3D-M behavior recognition algorithm with simplified model and high accuracy is selected to use in the vehicle taillight status judgment for FCW system.Because FCW system has high requirements on the accuracy of signal lamp state judgment,this paper proposes an X3D-M-SimAM algorithm combining with the non-parametric attention mechanism.Not only the network parameters and computation amount of X3D-M behavior recognition algorithm are reduced,but also the accuracy of taillight behavior recognition is improved.Experimental results show that the FCW system designed in this paper has a good measuring effect of the vehicle moving in the same direction,and the measuring error can be controlled within 6%,which meets the requirements of national standard GB/T21555-2007.Meanwhile,in the collected taillight data set,the accuracy of X3D-M-SimAM behavior recognition algorithm for taillight state status judgment of vehicles can reach 93.09%.The FCW system designed in this paper has high reliability and can provide effective information in the driving environment ahead of the vehicle for the ADAS system to assist the driver to effectively intervene in the vehicle driving state.