Research On The Recognition Of Human-vehicle Collision Based On Video Image

With the rapid development of road traffic in China,more and more vehicles travel on the road,and traffic accidents occur frequently,which seriously threaten the safety of people's life and property.However,the detection and processing of traffic accidents in our country is still through the way of telephone alarm and manual monitoring,and the efficiency is relatively low.Therefore,using video image processing technology to realize intelligent detection of traffic accidents has become a hot research topic.But at present,there are many problems in the research of automatic discrimination system for traffic accidents,such as inaccurate target detection,poor tracking effect,low accuracy rate of conflict discrimination,and so on.Moreover,most of the research contents are aimed at collision recognition between vehicles.There are few studies on human-vehicle collision.The casualty rate of a person after being hit by a vehicle is very high,and it is difficult to be detected and identified because of the severe deformation of the person.Therefore,in view of these problems,this paper proposes the human-vehicle collision accident recognition technology based on the video image.The main research contents include moving target detection,tracking and collision.(1)Moving target detection is the first step in human-vehicle collision recognition.By analyzing different moving target detection methods and background modeling methods,the moving target detection algorithm based on Gaussian Mixture Model is selected to achieve the optimal moving target detection effect.After extracting the foreground target region,the detection result is analyzed,and the shadow detection and removal based on multi-feature fusion is proposed.The color feature,texture feature and edge feature of the target are effectively combined to remove the shadow in the foreground target area,so as to optimize the detection results of moving targets.Finally,the validity of the algorithm is verified by experimental analysis.(2)Then the target tracking will be carried out.By analyzing different movingtarget tracking methods,the target tracking method based on kalman filter is selected,establishing kalman filter to predict the target,matching the target through the changing characteristics of centroid distance and area size of the target,and selecting the best matching value as the target feature information to complete the tracking of the moving target.Because of the problem of tracking failure caused by target occlusion in target tracking,this paper proposes a method for judging and processing occlusion of moving target,using the predicted value of kalman filter instead of the observed value to predict and track,so as to achieve anti-occlusion target tracking.Finally,the effectiveness of the algorithm is proved by experiments.(3)After the target tracking obtains the target information,the collision accident is recognized.The collision detection model of two monitoring perspectives in the region to be detected is established.According to the results of target tracking,the collision detection based on directed bounding box is carried out for videos of two perspectives respectively.When the collision of the same target is detected at the same time,the collision accident is considered to have occurred.Finally,according to the task requirements,the development of the identification system for human-vehicle collision accident is completed.Through the system test,it is proved that the identification system in this paper can effectively identify the human-vehicle collision accident.