Research And Implementation Of Vehicle Forward Collision Warning System Based On Machine Vision

With the rapid development of intelligent technology,vehicle intelligence has become one of the most popular research directions.The effective combination of advanced vehicle assistance systems and the Internet of Vehicles can provide drivers with a safe driving environment,reduce the probability of traffic accidents,and improve driving safety.Among them,the vehicle forward collision warning is one of the key technologies in the vehicle’s advanced assistance system.In the front vehicle collision warning method based on traditional vision,due to the low target detection accuracy,the inability to effectively measure multiple targets and the inaccurate collision risk assessment strategy,the entire system has a large warning error.In view of the above problems,the main The research work is as follows:(1)Aiming at the requirements of target detection accuracy and real-time performance,a forward vehicle and pedestrian target detection method based on deep learning is studied.Optimize the YOLO V3 network model to improve the detection accuracy.By adding a spatial pyramid pooling structure,using GIo U to optimize the network structure and loss function respectively,and create a vehicle road data set for training,and get the application in the forward collision warning scene Target detection network model.(2)Aiming at the system’s requirements for multi-target ranging,a target tracking method combining deep learning and tracking algorithms is studied.The target detected by YOLO V3 is input into the Deep Sort target tracking algorithm to obtain the target’s trajectory,and the monocular ranging algorithm is studied,the cause of the monocular ranging error is analyzed,and the target tracking result is compared with the monocular ranging Combined,real-time calculation of the distance between different targets of the vehicle in front and pedestrians.(3)Aiming at the problem of inaccurate collision warning,first formulate the calculation method of minimum braking safety distance and relative collision time.Then,the minimum braking safety distance and the relative collision time are combined to divide the road dangerous area,and the warning information is divided into three levels: prompt,warning,and warning for warning,which improves the accuracy of the system’s collision warning.(4)A low-cost hardware terminal system is designed,the deep learning network model is transplanted to the embedded terminal,the multi-process software architecture is adopted,the collision warning software and the visual upper computer software are designed to realize the collision warning of the vehicles and pedestrians in front.Finally,combined with the mobile cellular network to report early warning information in real time,and the video information to the Internet of Vehicles operation and management platform,the remote safety monitoring function of the vehicle is realized.The experimental results show that compared with the original model,the optimized YOLO V3 has improved the accuracy of vehicle and pedestrian recognition by 1.2% and2.3%,respectively.At the same time,the running frame rate on the terminal is 20 frames per second,and the real-time performance is better.In addition,the distance measurement error for the preceding vehicle and pedestrian is within 7%,which meets the distance measurement requirements of the collision warning system.In the end,the designed system realizes early warning of dangerous collisions,and combines with the Internet of Vehicles operation and management platform to effectively complete the remote safety supervision function of vehicles.