Research On Methods Of Behavior Recognition Using Feature Fusion

As a branch of artificial intelligence,action recognition is widely concerned by many researchers and research institutions because of its potential application value.The fundamental purpose of behavior recognition is to let machines analyzing who is in the video,where are they and what they are doing in the video automatically by using learning algorithms.The research results of behavior recognition have played a great social and economic benefits in many aspects,such as intelligent surveillance system,video based retrieval,human-machine interaction system,supplementary medical treatment,etc.Based on the theoretical research and practical application of behavior recognition,based on the review of the trend of behavior recognition and the research status at home and abroad,the following works are mainly done.(1)reviewed the development process of behavior recognition.According to the feature extraction method,the behavior recognition method was divided into the traditional manual feature extraction method and the deep web learning feature method.Combing the most commonly used data sets since 2001,and counting the citation times of each dataset in recent three years,side analysis confirms the development trend of behavior recognition.Some representative behavior recognition methods are analyzed and compared on several classic data sets,and the future trend of behavior recognition is discussed.(2)in view of the shortage of single feature descriptors in describing video spatial information,a fusion method that maps location information to visual features is proposed.This method is through the sampling point distribution of residual error characteristics and group characteristics and the two features are clustered according to the clustering principle of location information,and location information will be transmitted to the visual features,and then the visual features and mapping of the residual characteristics and group characteristics of location information fusion for video representation.In order to improve the performance of the traditional VLAD coding method,purposed an improved VLAD coding method.The experimental results on two large data sets of UCF101 and HMDB51 show that the proposed algorithm is improved.The experimental results on two large data sets of UCF101 and HMDB51 show that the algorithm proposed in this paper has a certain improvement in recognition accuracy.(3)In order to improve the characteristics of the dense trajectory feature,the ability of the video characterization is strong,but the data dimension is high and the computation is large.This paper purposed to divide the improved dense trajectory into four sub features according to the classification of descriptors.The problem of processing high dimensional data is converted to the problem of processing four low dimensional data.In order to solve the redundancy problem in the feature data,this paper also proposes a reduced dimension whitening preprocessing for the feature data before coding to remove the correlation between the data.The experimental results on UCF101 and HMDB51 datasets show that the algorithm is better.