Research On Behavior Recognition Algorithm Based On Optimization Modeling Of Low-level Feature

Behavior recognition and detection is an important field in computer vision.With the increasing demand for video surveillance applications and the continuous development and maturity of Internet technology,more and more applications involve the automatic recognition of video events.However,human behaviors are complex and diverse,which results in no feature models can be used universally.In recent years,the bag of visual words(BOVW)model based on low-level feature is widely used in the research of behavior recognition algorithm.However,the model ignores the weight of each visual word,and secondly it does not consider the spatial and temporal distribution of STIPs,which degrades the recognition accuracy.In this paper,we propose three new algorithms to solve the above problems.The main contents are listing as follows:(1)Term frequency–inverse document frequency(TF-IDF)method was used to determine the weight of every visual word in the traditional BOVW histogram.The BOVW model in computer vision is evolved from the BOW model for text categorization,therefore,the TF-IDF is feasible in BOVW model.First,when generating a visual dictionary through the clustering algorithm,the words in the dictionary are not distributed evenly,therefore,the distinguish ability of word is different.The importance of visual word is determined according to the its proportion in the bag of words and the BOVW histogram.(2)The STIPs mutual information(STIPs MI)algorithm based on three dimensional Co-occurrence matrix is proposed to describe the spatial-temporal relationship of interest points between different visual words.Firstly,the concept of two dimensional interest point co-occurrence matrix in the image is extended to three dimensional spatial-temporal interest points to generate co-occurrence matrix that describes the correlation between different interest points.Then the concept of STIPs mutual information is proposed toreduce the dimension of the co-occurrence matrix while maintaining the spatial-temporal information between different STIPs,and the STIPs mutual information descriptor is generated.(3)The concept of Annular STIPs histogram and spatial-temporal distribution entropy(STDE)that explore global distribution of STIPs is introduced for mining the spatio-temporal information of interest points within each word.Firstly,the algorithm constructs the annular STIPs histogram of each visual word,and then describes the spatial-temporal distribution of each visual word by STDE.Finally,optimized TF-IDF histogram,STIPs mutual information and spatial-temporal distribution entropy are fused as the descriptor of the video sequence,and the support vector machine SVM is used as a classifier.This paper is supported by National Natural Science Foundation of China(61472196 and 61672305),and the experiment is completed on Matlab platform;the proposed method is evaluated on two popular human action datasets: the KTH dataset and the UCF sports dataset.Experiment results confirm the validity of our approach and better than BOVW model and other mainstream methods.