Research On Behavior Recognition Algorithm Based On Multi Feature Fusion

In the information society, there is a lot of information produced every day. Information overload will cause information redundancy, which leads to decreased efficiency. People want to have a smart information carrier that can help them to perceive things intelligently. Behavior recognition is an important application in the field of computer vision. It can alleviate this problem to a certain extent. At present, behavior recognition has broad application prospects and profound economic value in intelligent monitoring system, intelligent robot, game entertainment, and virtual reality. As the human behavior recognition technology is a challenging problem and has great application value, it stimulates more interest in a growing number of researchers. However, most of the behavior recognition methods are designed for a simple action. They have good recognition results only on a specific research direction.In this paper, we study and proposed a behavior recognition method for simple human action and test it on the Weizmann and KTH datasets, respectively. For motion detection, the moving region is obtained by using a frame difference method. Typical feature point detection algorithm, including SIFT, SURF, and ORBS are discussed in detail. Their performance in the case of rotation, shift and size changing are compared. As a result of the comparison, the SURF algorithm is adopted to implement feature point detection. It has been proved that the human body contour and their speed information reflect human’s behavior in the video. HOG descriptor has a distinct advantage in describing the edge and contour information. Meanwhile, the HOF descriptor can reflect the velocity information of moving bodies. Therefore, the HOG descriptor and the HOF descriptor are selected to make a strong descriptor for feature fusion by series connection. The descriptor is regarded as a word bag model. It is clustered into 4000 classes by using the K-means clustering method and the cluster centers are obtained. Through the calculation of Euclidean distance between other video descriptors and the dictionary, the descriptor is mapped to a feature histogram to form the feature vector. In terms of classification, we use SVM as the classifier in this work. In experimental section, several tests are performed respectively in different data sets. The average value of the test results is selected as the experimental result of this work.Experimental results of the proposed method are compared with different methods that are published on international journals and conferences. The comparison result proves that the proposed method has advantages of effectiveness and generality.