Abstract: The cost of monitoring human actions is very high through off-line learning.The authors propose a system to solve the problem.The system has the capacity of real-time online learning and monitoring human actions.And the system consists of three weight factors, distribution of trajectory center difference, body tilt angle and ratio of height versus width.The three indexes are weighted and averaged, the quantitative index for behavior action is obtained, then the human behavior is judged according to the value of action: when the values greater than 2, it is considered normal, less than 1.5, the warning is needed while less than 1 it is considered abnormal.The result of experiment shows that this approach is superior in performance on real-time and the shortwoming of the high cost is overcome.