Vehicle Hazard Identification Based On Machine Vision

With the increase in vehicle ownership,driving safety is becoming a serious problem.More and more academic and industrial circles have been studying the onboard Advanced Driving Assistance System（ADAS）from the perspective of vehicle active safety to alleviate the current serious driving safety problem.Currently in the automobile stock market,most vehicles still do not equip the ADAS provided by the original factory.Therefore,it is of great practical and theoretical significance to carry out the research on the identification of the vehicle’s hazardous external environment based on the rear-loading equipment.Based on this,the appropriate on-board embedded equipment is selected to support monocular vision to perceive the external environment of the vehicle,and the two key points of the vehicle hazard identification system based on machine vision,which are vehicle forward collision warning system and traffic light detection and signal recognition system,have been studied.The specific research contents of the paper are as follows:（1）Aiming at the problems of low detection accuracy and poor recognition performance of small-scale targets in traditional vehicle and pedestrian detection methods,a vehicle and pedestrian detection method based on improved YOLOv4-Tiny is proposed.On the basis of YOLOv4-Tiny,the 4-fold down sampling feature layer was added for feature fusion,the PANet structure was used to perform bidirectional fusion for the deep and shallow feature layers from the backbone,and the detection head for small targets was added.The experimental results show that the mean average precision of the improved method for vehicle and pedestrian detection on KITTY dataset has reached 85.93%,which is 24.45% higher than that of YOLOv4-Tiny,and the running speed on the embedded platform Jetson Nano can reach 14 frames per second.（2）Focus on the requirement of vehicle forward collision warning system for target ranging,the calculation formula of monocular ranging is deduced based on the geometric projection model,and the distance calculation of the target in front of the vehicle is realized.On the basis of ranging,a vehicle forward collision warning method based on the minimum braking safety distance and the Time to Collision model is proposed,and the warning information is divided into prompt state,warning state and dangerous state to inform the driver.（3）In view of the problems that the scale of traffic light is small and the computational power demand of the current detection algorithm is too large to match the on-board embedded system,a lightweight model based on YOLOv5 s for traffic light detection and signal recognition is proposed.The experimental results show that compared with directly using YOLOv5 s to identify traffic lights,the average recognition accuracy of the proposed algorithm is improved by 45.4%,while the amount of parameters is reduced by 73.3%,the calculation quantity is reduced by 79.2%,and the running speed on the vehicle embedded platform Jetson Nano is increased by 87.5%,which can reach 15 frames per second.This research results can provide theoretical foundation and methodological reference for the development and popularization of ADAS.