Algorithm Research On Vision-Based Human Motion Analysis

Vision-based human motion analysis is one of the most active research fields in computer vision. It has many promising applications such as intelligent surveillance, advanced user interfaces and virtual reality. Vision-based human motion analysis aims at detecting, tracking and identifying people from video sequences, and furthermore understanding and describing human behariors. In this paper, some researches have been done on the algorithms involved in motion object detection, object tracking and human motion recognition.An algorithm of background subtraction based on Gaussian mixture model is implemented. And a learning algorithm of background model construction based on adaptive learning rate is used. This method converges faster and detects moving object more efficiently, especially under the condition of objects staying in the initial scene.An object tracking algorithm based on the Particle Fileter theory is implemented in this paper. The color histogram is integrated into the tracking algorithm and this can reduce the number of particles needed and the computing load. Furthermore, the algorithm exhibits robuster in the situation that the object is partially occluded, rotated and shape distorted.A vision-based human motion recognition method is introduced in this paper. A modified shape context method named area-based shape context is used to describe the shape. Then continuous hidden markov models are built to model the human motions, and Bayesian criterion is used to classify them. The method can classify five kinds of human motions.The proposed novel ideas in this paper are as follows:1. The adaptive learning rate based background model construction algorithm improved the conventional GMM in two aspects. In the model initializing process, a learning algorithm is used to construct a background model. Different learning mechanisms are respectively used in model initialization and model updating process. The adaptive learning rate is used in model updating in order to get a fast convergence.2. This paper puts forward area-based shape context technique as an expression of human shape. This method pays more attention on the whole human body area, instead of detail information. It lowers the requirement of the integrality and the precision of the human shape and can work well when the image is not clear enough. 3. Continuous hidden markov models are built to construct human motion models. The best state number estimation method is proposed to find the best state number for each HMM.