Research On Human Behavior Recognition Method Based On Space-time Interest Points And Bag Of Words Model

Human behavior recognition technology is an important research field in the field of computer vision.It has a wide range of applications prospects in many aspects such as human-computer interaction,video surveillance,virtual reality and so on.The human behavior recognition method based on spatio-temporal interest points building bag-of-words model has the advantages of low computational complexity,insensitivity to background noise and strong robustness,In recent years,it has attracted a great deal of attention from researchers in human behavior recognition.With the joint efforts of researchers,Although the algorithm has made significant progress in human behavior recognition research project,there are still some problems such as poor visual vocabulary discrimination and large hard quantization errors in feature coding.In view of this,this paper puts forward the corresponding improved algorithms based on the further study of the bag-of-words model.The specific research work is summarized as follows:1)An algorithm based on hierarchical clustering is proposed to form more representative and distinguishable visual words,which makes the intra-class similarity high and the inter-class difference high when video is represented.Combining with the feature fusion of the video expression level can effectively represent the video.Experimental results show that the proposed algorithm has obvious improvement in both operational efficiency and recognition rate.2)An VLAD encoding method of enhanced local descriptor aggregation vector is proposed.The algorithm can effectively eliminate the existing abnormal feature descriptors and adopts the soft quantization strategy for VLAD coding through local linear constrained coding coefficients,taking full account of the spatial distribution between visual words and feature descriptors.Experimental results show that the proposed algorithm improves the performance of video feature coding.3)A method of human behavior recognition based on BP neural network classifier is proposed.The research on the initialization of weights and the selection of activation function and the use of Dropout technology prevent overfitting effectively and optimize BP neural network model,which has good performance in classification and recognition.The experimental results show that the proposed method has better recognition performance than the commonly used classification methods such as KNN,Bayesian classifier and SVM.