Study Of Video Human Action Recognition Based On Local Spatio-temporal Features

Nowdays，in the field of computer vision and multimedia systems, the scientificcommunity has given particular attention to the automatic recognition and interpretation ofhuman behaviours. The applications of such research are in the context of classification andretrieval of multimedia contents, humancomputer interaction and surveillance systems.Recent works have been focused on describing human actions using local descriptiontechniques of moving patterns. The description is achieved through algorithms that detect anddescribe features who are named ‘space-time interest points’ in the spatio-temporaldomain.Representation based on local spatio-temporal feature has very good stability even ifsome action is changing of scale or orientation or illumination，therefor，the focus of thispaper researches video human action recognition based on local spatio-temporal feature.Firstly,this paper analyzes the basic traditional principles of video human actionrecognition system. According to various modules in the system,the system is divided intofeature extraction and model matching.In order to reduce the influence of the shortcomings ofcommon feature detector algorithm,this paper puts forward to a new feature extraction method.Then,in the phase of model matching, traditionally,the BoW model has much morereconstruction error due to the more restrictive constraint，this paper puts forward torepresentation before recognition using sparse coding instead of vector quantization,andcombines space pyramid with mapping three orthogonal planes (TOP) under max poolingmodel.At last, this paper designs experiments which divids video database into training videoand testing video by the ratio of7:3,which verifies the effctiveness of this system.In addition,this paper studies video human action recognition based on instance-to-class(I2C) distance.And it puts forward to using Naive Bayes Nearest Neighbor algorithm (NBNN)and Local Naive Bayes Nearest Neighbor algorithm (LNBNN) to classify test videos directly.Through experiment simulation, video human action recognition based on I2C distance has amore obvious performance improvement compared to the traditional SVM classificationwhich is based on instance-to-instance (I2I) distance.