Research On Video-based Vehicle Collision Detection Method

Traffic accidents caused by vehicle collisions have severely threatened the safety of people's lives and property,and have been widespread concerned by people.Many efforts have been tried to develop an automatic detection method via target tracking techniques by a lot of research teams,but the current video-based vehicle collision detection algorithms have many problems,such as the low accuracy rate for target detection,tracking and identification.Therefore,this thesis presents a vehicle collision detection algorithm based on motion interaction field.And then the thesis improves the algorithm with the Exponential Smoothing algorithm.The main work is as follows:Firstly,the thesis introduced the background and significance of our research,and then the thesis reviewed the study of the related content at home and abroad of vehicle collision detection techniques.Secondly,the thesis introduced the common algorithms of extracting the moving target information in the traffic video stream.By analyzing the application of different algorithms and the advantages and disadvantages of different algorithms,the thesis selected the polynomial-based dense optical flow algorithm to extract the motion information of each pixel of the image frame.Through the algorithm programming,we obtained and stored the speed,direction and other information of each pixel.Through experimental analysis,the algorithm achieved a good detection results.Thirdly,we have introduced the common algorithms of Traffic Accident Detection.By studying the theory of the hydrodynamic characteristics of two objects approaching to each other in viscous flow,we constructed the traffic model based on Motion Interaction Field.The experimental results showed that the algorithm can judge whether the vehicle is in a normal driving state,and can identify and locate the vehicle collision phenomenon in the image frame,and can judge the start time of the vehicle collision accident in the video.Then,the thesis analyzed the reason why the exact motion interaction field can not be constructed when the vehicle stopped moving after vehicles crashed.The thesis improved the traffic model based on the motion interaction field by the exponential smoothing method,which can use the abnormality value of the collision start frame to predict the abnormality value of the future image sequence.The experimental results showed that the improved algorithm can accurately judge whether there is a vehicle collision phenomenon in the image frame even if the vehicle in the image frame stopped moving due to the collision.Finally,the thesis analyzed the proposed algorithm under normal traffic and abnormal traffic,it proved that the algorithm achieved a good detection results.And it demonstrated that our algorithm can identify and locate the vehicle collision phenomenon in the video with high accuracy rate by comparing the algorithm with other existing algorithms under the 8-segment real traffic monitoring video.