A Vehicle Abnormal Event Detection Method Based On Compressed Domain

At present,the sampling and investigation of traffic accidents are mainly carried out by means of manual inspection,which is time consuming and laborious.In recent years,the development of computer vision technology for the detection of abnormal vehicle events provide a lot of solutions which are based on the pixel domain method.However,due to the high resolution of the currently monitored video and the decoding characteristics of the pixel domain method,making the traditional pixel domain method is difficult to meet the real-time requirements,while compressed domain methods have a great advantage in real-time requirements.Through the study of the current latest compressed domain coding standard HEVC and the analysis of a large number of traffic accident videos,it is found that the partition mode and motion vector information contained in the HEVC can be viewed as course analysis of target boundary and optical flow,which may be useful for traffic event detection.The current research in compressed domain is mainly for the segmentation and tracking of moving targets.The main purpose of these algorithms pursue high segmentation precision which makes a high computational complexity.On the basis of HEVC coding standard and coding characteristics,this paper propose a motion intensity value algorithm based on the motion vector and block partition information.And on this basis,a multi-parameter fuzzy logic algorithm is proposed to detect abnormal vehicle events.The main work and innovations of this paper include:(1)As the motion vector and block partition information in HEVC has a high correlation with video content,this paper propose a moving target detection method based on the motion intensity value.Firstly,the block partition and the corresponding motion vector information are extracted from the HEVC compressed stream,the extracted motion vectors are proceeding cumulative iteration in inter frames and median filtering to get an orderly and smoothly motion vector field.Secondly,the motion intensity values of each largest coding unit are calculated in conjunction with the block partition information.Finally,through the set of adaptive threshold value,the coding units with smaller motion intensity value can be filtered.So that the detection of moving targets can be achieved based on the motion intensity value.(2)In the current pixel domain method,the traffic accident detection algorithms are mainly based on the centroid coordinate distance difference and the time difference,which have high computational complexity and poor reliability.With the analysis of the difference of area and the motion intensity value,this paper proposes a multi-parameter fuzzy logic algorithm to detect vehicle abnormal events.Firstly,the breadth-first search algorithm is used to construct the connectivity region for the largest coding units contained motion intensity value,and then the area of the connected area and its corresponding motion intensity value are extracted.Second,according to the area gradient information to distinguish between different models and the traffic conflict scenes.Finally,the motion intensity value,area gradient information and the deviation value from average of motion intensity value are constituted as the input parameters of fuzzy logic theory.According to the vehicle movement characteristics of the membership function to determine whether the vehicle abnormal events happened in the current frame.Through the experiments of 25 groups of traffic accident videos,22 of which can be accurately identified and the detection rate is 88%.