Research On Moving Targets Classification Algorithm In Surveillance Video

Moving target classification is an important part in video analysis. As the securit y problems becoming increasingly serious, video surveillance has become an essential way to protect safety. It is necessary to make research on how to extract useful information from vast of videos. In this paper, based on moving objects detecting, tracking, through series of processing to surveillance video, moving targets category information can be acquired, which provide a technical basis to event analysis and behavior understanding in surveillance videos.Firstly, a framework for moving target classification is established. On the basis of the framework, we pay a lot of attention to key technologies such as feature extraction and description, classification algorithm in moving targets. To verify the effectiveness of the framework, experiments are carried out. The main contributions in this paper are as follows:A framework of moving targets classification is proposed. The process of moving target classification can be divided into four levels: object detection and tracking, feature extraction, characterization and classification. A classification framework based on local features and texture features is proposed, which is in accordance with the underlying frameworks. Making analysis to surveillance video expand from low level features to high-level semantics, the semantics acquired can make a connection between middle-level features and high-level semantic descriptions of surveillance video, which lay the foundation of event analysis and behavior understanding in surveillance videos.Secondly, a target description algorithm based on uniform-rotation invariant local binary pattern(riu-LBP) texture feature is proposed in order to overcome the shortcomings of shape features, which performs well in changes on deformation, rotation, angle, scale.Third, a characterization algorithm for targets is proposed based on bag of words model. The procedures are as follows: All local features in the target area are extracted; Then apply bag of words model into multi-dimensional vectors to get a one-dimension vector. It can improve the accuracy of moving targets classification, as well as the efficiency.Fourth, a method for moving target classification is proposed.The selection of classification method is very important to classification framework. The methods in this paper are single-core SVM method, K Nearest Neighborhood classification. In order to improve classification accuracy, the best classification model should be obtained through comparison.Last but not least, an experiment is carried out to verify the effectiveness of the proposed classification method for moving targets, which lays the foundation for the follow-up researches in the future.