Research On The Key Issues Of Intelligent Behavior Perception Based On Video Content

With the development of computer technology and the wide application of computer vision principle, computer understanding human motion and behavior has become an important development direction of high-level computer vision technology. Based on the summary of previous human behavior feature extraction and recognition algorithm, around the human behavior perception problem in the real world two-dimensional video, starting from the extraction of human body contour and human behavior characteristic, this paper is focuses on shape boundary feature description algorithm and Fourier descriptor, and the matching application of extracted features in Dynamic Time Warping problem. Combined with the environmental context, a continuous human behavior perception algorithm framework is put forward. The framework forms the basis of human behavior analysis algorithm by tracking and recognizing many human behavior video. The main contents are as follow:(1) Study on the basic theory of image target detection and segmentation, including Canny detection operator and mixed Gaussian background modeling method and its principle and process and characteristics. Detect the human body action outline of discontinuity point on the basis of the mixture of the two algorithms.(2) Study on the different human behavior feature extraction and representation, including the comparison of the accuracy in centroid distance, Fourier descriptor, Hu moments and skeleton properties. Also study the Theoretical basis and algorithm process of Dynamic Time Warping. For the matching standard of video sequence in the warping process, analyze the advantages and disadvantages of four kinds of similarity measure algorithm and choose one of them as body posture matching criterion.(3) The human behavior perception based on Dynamic Time Warping implemented. Focuses on the selection in shape feature descriptor and similarity matching criteria when recognize the human behavior. Finally implement behavior comprehension steps to the bottom recognition results combining with the context. With the video behavior public database, test the algorithm framework and obtain good behavior perception results despite there is low accuracy and weak robustness.(4) Propose an algorithm similar to the method of minimum spanning tree to fill the outline for more precise tracking result from human body segmenting an open continuously outline. In order to compensate the problem for matching distance easily affected by scale when the feature vector used in matching calculation, this paper ues adaptive weighted fusion method for Fourier descriptor and center distance which calculating the feature vector' slope difference instead of absolute difference and make the behavior recognition accuracy improves.