Research Of Vision-Based Human Behavior Recognition

In recent years, video-based analysis of human behavior is a hot research field of computer vision. It has broad application prospects in intelligent video surveillance, sports and entertainment motion analysis, human-computer interaction and virtual reality, etc.The thesis focuses on issues related to human behavior based on video analysis from the following three parts:the moving target detection, feature extraction and behavior recognition.In the moving object detection, the thesis uses Gaussian mixture model to establish background model and obtains moving foreground by background subtraction method. Then Canny operator is used to make global edge extraction. The movement target and the contours of the binary image obtained from the two above methods are calculated using logical "or" operation and morphological processing.In the feature extraction, the thesis aims to find a good feature descriptor, which not only describes motion information but also contains the body shape features. Hu moments can be used to obtain the information of the body contour. The histogram of oriented gradient (Hog) is used to orient the local information of structure and appearance. In the thesis, these two features are combined to get a better feature descriptor. This feature descriptor contains not only the body shape features, as well as the motion area information. It is a good feature descriptor and has provided a good foundation for behavior classification and recognition of human actions in the following process.In recognition of human behavior, the thesis employs the radial basis function of multi-class support vector machines for classification and identification of human behavior. K-cv cross-validation method is used to validate the accuracy of the method.The thesis uses Microsoft Visual Studio 2010 platform and programs based on OpenCV computer vision library. The accuracy and real-time performance of the algorithm proposed in the thesis are tested using multi-segment videos in the Weizman video database. Experimental results indicate that algorithm proposed in the thesis can detect human motion in real-time and recognize the human behavior accurately. Compared with the behavior recognition method based on a single feature Hog or Hu moment, feature extraction algorithm of this paper is slightly higher than the cost of Hu moment feature in time but similar with the Hog feature. In the classification performance, this algorithm has a broader scope and higher average recognition rate. In further it is more effective to identify a specific human behavior under certain scenarios.