Research On Video Abnormal Event Detection Technology In Crowded Scenes

In recent years, the intelligent video surveillance technology (IVS), driving thebooming video surveillance market, has made a wide range of important applications inall aspects of life. As one of the key research topics of IVS, video abnormal eventdetection is to detect the small amount of abnormal events that do not match normalevents automatically and alarm timely from a large number of surveillance video by thecomputer vision technology. Detecting abnormal events especially in crowded scenesremains challenging due to the large number of occlusion of research, as well as thediversity of events defined by various applications.In this paper, we introduce the key technical methods in video abnormal eventdetection; especially discuss the event motion feature representation and learning modelin detail. The main work and innovations are as follows:(1) We summarize three types of main motion feature representations, includinglocal descriptors, optical flow and the dynamic texture. To begin with, we give a briefdescription of theory and algorithm of two popular optical flow method, namelyHorn-Schunck method and Black-Anandan method, and then conclude the motionfeature in localized region by multi-scale histogram of optical flow (MHOF) algorithm.Furthermore, we propose a novel dynamic-texture for event motion representation,called Structural Multi-scale Motion Interrelated Patterns (SMSMIP). SMSMIPcombines both original motion patterns and their structural spatio-temporal information,which effectively represents localized events by different resolutions of motion patterns.In addition, the histogram down-sampling method can greatly reduce the dimension ofseries multi-scale motion features firstly. Then the subspace learning method, includingprincipal component analysis and whitening method, is adopted to calculate the finaltransformation features.(2) We summarize two types of learning models for motion features, including theGaussian mixture model based on likelihood estimation, and the model of vectorquantization, sparse coding and locality-constrained linear coding (LLC) based on errorreconstruction. The LLC model is the first time applied in the abnormal event detection field in this work, then improved by the adapted nearest neighbor codebook method tocalculate the reconstruction error of sample more stably and precisely, which brings lowcomputational complexity and high precision. Meanwhile, the proposed model can belearned online by updating the parameters incrementally in order to adapt to changes inthe video stream data.In order to compare and evaluate the effectiveness of the proposed models, thispaper analyzes how the main parameters in the model effect the detection rate andcomputation time at first, the combinations of different motion features and learningmodels are then tested in three different types of popular datasets. The quantitativecomparisons show that the results on the UCSD Ped1and UMN Dataset outperformrecent state-of-the-art approaches. SMSMIP can represent the motion patterns of eventsbetter compared to MHOF; the structural spatio-temporal information added in motionrepresentation helps increasing the anomalies detection rate by average almost5%. Alsothe promising results on the Subway Exit dataset demonstrate the effectiveness androbustness of our method.