Moving Human Behavior Recognition In Natural Scene

With the progress and development of society,people pay more and more attention to security issues.Therefore,the demand of video surveillance system is increasing rapidly.Video based human behavior recognition is one of the important branches of video surveillance system,which has wide application prospects.However,in the actual scene,there are still many problems in human behavior recognition because of the problems of visual angle,occlusion,light and so on.This paper mainly studies the moving human recognition based on video in real scene.The work is divided into four aspects in this paper:moving target detection,human behavior feature extraction,feature coding and feature fusion.The major works are summarized as following:(1)The optical flow method is the most widely used algorithm for moving target detection,but its biggest disadvantage is the high complexity of time.In this paper,an improved optical flow algorithm is proposed in order to make up the shortcomings of the Horn-Schunk optical flow.On the one hand,the frame difference method is used to reduce the time consumed by the iterative process.On the other hand,the gradient is optimized to further improve the precision of moving target detection.(2)Different feature descriptors depict the motion of human behavior from different point of view.In the Dense Trajectory algorithm(DT),the features extracted along the trajectory include the trajectory features,HOG,HOF,and MBH.On the basis of DT algorithm,the influence of camera motion is considered in the improved Dense Trajectory algorithm(iDT).The iDT algorithm uses the optical flow between two adjacent frames and the matching of the key points of the SURF to estimate the motion of the camera and remove the background interference.In this paper,the feature extraction parts of the DT algorithm and the iDT algorithm are studied,which lays the foundation for the follow-up processing of multiple features.(3)Bag of Word and Fisher Vector are the commonly used feature coding methods,but they both have disadvantages.In order to reduce the complexity of time and space at the same time,the human behavior recognition algorithm based on Kmeans++ and VLAD is proposed in this paper.The Kmeans++ is used to cluster the various features to ensure a more accurate clustering center.In order to improve the progress of experiments,we take advantage of the rapidity of GPU and complete the VLAD feature encoding of behavior database UCF50 and Hollywood2.(4)It is very limited to use a single feature to recognize human behavior in video.Therefore,an effective feature fusion method is needed.Aiming at the problem of directly connecting multiple features and neglecting each feature to affect the recognition results,human behavior recognition based on multi-feature weighted fusion is proposed in this paper.The feature weights are allocated by normalization to ensure that the more significant the weight value,the more significant distribution it is.Experiments in behavioral database UCF50 and HMDB51 show that Kmeans++ and VLAD human behavior recognition algorithm based on multi-feature weighted fusion recognition method can further improve the recognition rate of human behavior in video.