Research On Road Target Detection And Tracking Technology For Vehicle Collision Warning

Vehicle collision warning is an indispensable technology in the field of intelligent driving.It can provide early warning of possible collisions by acquiring the relative speed and position information of the target in front of the vehicle.Traditional collision warning systems use radar for information acquisition,but this method has disadvantages such as complex algorithms,high cost,and low accuracy.In recent years,the advancement of image technology has made the camera information acquisition program more feasible.In order to realize the vehicle collision warning based on image technology,this paper studies the real-time road target detection and tracking technology according to the requirements of the GB/T 33577 standard for the collision warning system.The main research work is as follows:(1)Research on road target detection technology.The principle of target detection technology is introduced,and after comparing several detection algorithms,the YOLO algorithm that can better meet the technical requirements of collision warning is selected.Combined with the latest YOLOv5 s algorithm,the YOLOv5 s model of road target detection required by the vehicle collision warning technology is proposed and improved,and two methods that can effectively reduce the weight parameter of the convolutional neural network are explored.Two lightweight road target detection models,DS-YOLOv5 s and Ghost-YOLOv5 s,are established.(2)Trained and verified the established road target detection model.In order to meet the data set requirements for model training,the road targets in the BDD100 K data set were extracted,and different scenarios were selected to verify the model.The results show that both algorithms can meet the requirements of collision warning.The Ghost-YOLOv5 s model The real-time performance is the best.(3)Through the combination and improvement of the Ghost-YOLOv5 s model and the deep SORT algorithm,a road target tracking model is established,and a target association siamese network is built in combination with the cross-phase local network,and the targets are associated,using two scenes to match the road targets.The tracking model has been tested,and the results can better meet the requirements of the vehicle collision warning system.