Video Abnormal Events Detection Based On Low Level Feature

In recent years, with the occurrence of a variety of public safety, disaster monitoring, Social anti-terrorism and other types of events, surveillance video have become ubiquitous in our lives and the amount of surveillance video also showed explosive growth.The traditional way to manually monitor video information processing is not only ine?cient but also very easy to miss suspicious objects and abnormal events. For this reason, intelligent surveillance video analysis has become an increasingly urgent demand.How to use computers to make intelligent understanding of surveillance video content more e?cient has become a hot area of research recently.This paper focuses on the detection of abnormal events in videos:(1)A preprocessing algorithm of optical ?ow macroblock is proposed: for lowlevel features video, we do not limited to two traditional approaches——pixel-level feature as a separate whole or macroblock feature without preprocessing as a separate whole, differently, we ?ltrate high motion consistency pixels as a separate whole by the preprocessing of optical ?ow macroblock to expressed as the the form of a collection of optical ?ow vectors.(2)Propose a new statistical features model based on the optical ?ow macroblock: creatively introduce information entropy theory into statistical feature extraction process, the phase features of optical ?ow macroblocks have expressed very well by quantify the direction of the optical ?ow ?eld and calculate orderly coe?cient of optical?ow macroblocks. Then use these two features to build the statistical feature of the optical ?ow macroblock.(3)In the design of algorithm framework, we also made innovative contributions:in the choice of machine learning models for feature training, fully take the physical meaning of the statistical features into account, Gaussian mixture model was chosen to training the extracted statistical features of optical ?ow macroblock, making the trained model is more in line with the abnormality degree of abnormal optical ?ow macroblock.Then the extracted statistical features of test video dataset was ?ltered through Gaussian ?lter and make a determination of abnormal events becasuse Gaussian mixture model is more sensitive to continuous variable features.Four video dataset are chosen to test the proposed algorithm framework, UCSD video dataset, UMN video dataset, Subway video dataset and U-turn video dataset.The experimental results show that the proposed statistical features of optical ?ow macroblock could effectively represent the abnormal events of videos. The proposed algorithm framework not only has a strong ability to adapt to many scenes, and in each scene could have good results to detect abnormal events according to comparisons with other existing algorithms on quantitative indicators AUC and EER in four video dataset.The proposed algorithm, to some extent, solves video abnormal events detection with partial occlusion, at the same time, has better abnormal events localization in video abnormal events detection.