Video Unusual Event Detection Based On Loop HMM-LDA Model

Today, video surveillance system plays an irreplaceable role in security inspection of airportsand railway stations, urban security maintenance, traffic control, and national defense security. Butthe traditional video surveillance system which is almost analyzed by human beings no longer hasthe ability to deal with the massive video data which is increasing at a high rate, and even it cannotguarantee the reliability and efficiency. How to improve the intelligence, efficiency and accuracy ofvideo surveillance systems will be a big problem which is needed to be solved quickly. Based onprevious studies, this article, taking the video abnormal event detection as the main task, exploreshow to improve the intelligence, efficiency and accuracy of video abnormal event detection fromcomputer vision and image processing technology. The specific research work I have done is asfollowing：After studying feature extraction, subject feature extraction, and feature analyze andrecognition technology of video abnormal event detection, this paper proposes a detection methodbased on LDA-HMM for abnormal video event to solve the two questions of traditional detectionmethods. One is the lower detection rate led by probability dragging, such as the method based onstatistical method. The other is that traditional video abnormal event detection methods separatefeature extraction and feature analysis to process leading to low adaptability rate of single scene.The method in this paper firstly constructs LDA structure, iHMM structure by referencing theDDP-HMM structure and constructs a loop model which can be taken as up-loop, down-loop modeland the vertical direction can be regarded as the feedback neural network by simulating themechanism of the Elman feedback neural network. Secondly, completes the reference process of themodel according to the idea of re-weighting algorithm of loop tree. Thirdly, detects the abnormalevents through the model above.This paper designs detection process of video abnormal event detection and the GUI ofdetection system which is based on this detection method, In order to verify the detectionperformance of the detection method, this paper select four complex experiment scenes to carry outthis detection method and make an analysis to verify that the detection method is also effective formany scenarios. And then makes a comparison with other video abnormal detection methods indetection performance to verify this detection system have a relatively high detection rate.Traditional video abnormal event detection methods have the problems of low detection andadaptability rate. The loop LDA-HMM model in paper can realize the feedback between the spatialtopic feature and time evolution of states, which make the loop model adjust itself according to different scenes and adjust the extent of fitting between model and scene to the best. So it not onlysolves the problem of adaptability of single scene, but also improves the accuracy of detection rate.The simulation verifies that the test method is not only effective for muli-scene but has a relativelyhigher detection rate.