Human Action Recognition Based On Weighted Optical Flow And Concurrence Relationship Graph

The technology of human action recognition has become one of the most important technologies in the field of computer vision, and it has been widely used in the field of video surveillance, video and image retrieval, human-computer interaction currently. With the increasing researches in this field, the recognition rate has reached very high in simple scenes. However, the recognition rate with complex background will reduce greatly due to the camera shaking, light and angle changing, and the existence of occlusion.As the representative of the local features, Spatio-Temporal Interest Points(STIP) is becoming the most popular method in the stage of feature extraction at present. And it is robust to the interference produced by background. The goal in this thesis is to improve the recognition rate under complex scenes and we discuss the role of human action recognition in Intelligent Monitor. The main results of the thesis are listed as follows:1. A detailed discussion is made on several pivotal problems such as: estimate of motion region, pretreatment of image sequence, feature extraction and description, machine learning model and so on.2. A new feature of weighted optical flow by combining the optical flow field and Harris 3D detector is proposed. The integration of these two sources of motion information may provide the complementary motion information to improve the region of action estimation. With the feature of weighted optical flow, motion region and HOG/HOF descriptors are used to descript motion information, which can restrain interference produced by dynamic background in some degree.3. A concurrence relationship graph is given to eliminate the interference produced by dynamic background. The paper uses the context information of local STIPs to construct a concurrence relationship graph, through which the stable STIPs was selected in the realistic background. Then the Bag of Words model is re-established to generate the final feature vectors as the input SVM classification.The experimental results on KTH dataset and UCF Sports dataset demonstrate that the proposed approach has obvious effect on improving the recognition rates of realistic data.