This paper takes into account grid deep spiritual pupil grade-evaluation system and main ideas of hysteresis comparator, combined with the nearest neighbor classifier and SVM learning, which aims to establish an automatic Classification and Grading Warning System recognizing six common target abnormal behavior including invasion ,fast-moving , remnants, fleeing, fights and riots . This paper contributes at following three aspects: firstly, we propose a comprehensive system dealing with classification and warning of abnormal behavior. Secondly, before recognizing abnormal behavior the population density and the energy is characterized, which are input of the nearest neighbor classifier, achieving pre-classification of individual behavior and group behavior . Thirdly, this paper tends to achieve stable behavior warning by introducing hysteresis comparator, and this method has certain universal significance. In this paper, experiments are carried out on the standard library and video sets shoot by our own. Experimental results show that the system can achieve high warning and classification stability of the six abnormal behavior, which integrates self-analysis and analysis groups, detection and identification, classification and warning together.