Detection and identification of abnormal behavior is an important technology which can be applied in many cases. In this paper, an improved algorithm of positioning key points is proposed. It models the human skeleton parametrically based on the key points and converts the problem of human behavior trend analysis into the parametrical problem. It reduces the amount of calculation and storage in human behavior analysis greatly. The algorithm can deal with most body postures and the experiment shows that extracting results are more accurate and can be the base data for detection and identification of abnormal behavior for online examination.