Research On Multi-Person Behavior Recognition And Analysis Based On Sitting Posture

Visual analysis of human motion is currently one of the most active research topics in the field of computer vision. Human motion analysis aims to detect, track and identify people, and more generally, to understand human behaviors, from the video sequences involving humans. In recent years, although visual analysis of human motion received increasing attention from both academia and industry, many theoretical and technical problems remain open. Most research on human behavior analysis is based on single person, and simple actions, there is less research about multi-person and their mutual behaviors. Motion detection is the basic of the human behavior analysis and because of the noise in motion detection, such as varying luminance, the result of human behavior analysis always get poor effect. In this dissertation, a sitting posture recognition method was proposed based on PCA (Principal Component Analysis), and it was used in multi-person environment, then, mutual behavior recognition and analysis method based on the distance between two neighboring facial skin boundaries was devised.Firstly, the contour of motion object was obtained by background contrast attenuation method and it was filled by white pixels, then skin area of motion object was extracted using skin model. And furthermore, skin area image was normalized in order to make the skin area occupy the full image area, because skin area of motion object is only a small part of the current frame.Secondly, a sitting posture recognition method based on PCA was put forward. The training postures were used to construct the PCA subspaces and the grayscale image of skin of a new posture was projected onto the subspaces, and then sitting posture recognition was realized by cosine classifier. Facial rotation motion was further analyzed according to the change of the pixels of facial skin area in different video frames.Finally, mutual behavior analysis was realized by the distance of two neighboring facial skin area boundaries. The mutual behavior of two persons was regarded as abnormal when the distance of their skin area boundaries was less than the preset threshold, otherwise, each person area was segmented and PCA was employed again to recognize the behavior of each person.The experimental results show that the data providing method of sitting posture recognition in single person environment is reasonably robust in varying luminance and shadow, and the accuracy of behavior recognition by PCA is 85.15%. Furthermore the mutual behavior of two neighboring persons can be accurately analyzed and the behavior of each person can be exactly recognized.