Aiming at some limits in cross-border abnormal behavior recognition f or pedestrian in a video surveillance scenario at present,a novel approach to recognizing the cross-border abnormality has be en developed based on a spatial virtual wall.There are some rigid invariances and differences in the Haar-like features for the head. A cascaded classifier and dynamic tracking chain with the particle filtering are fused to track and locate the head as a target from the v ideo surveillance scene.Both three-dimensional plane equation for the top of the head and its corresponding spatial virtual wall in t he surveillance scene are constructed based on the height-invariance.The problem whether a pedestrian spans a warning line in a t wo-dimensional scene is transformed to the one whether he passes through the three-dimensional virtual warning wall.It is est imated efficiently if a pedestrian crosses the border or not.Some state-of-the-arts and experiments have been done in some video scen es with different contents to test the performance in the same conditions.Experimental results show that the proposed meth od is efficient and valid without any specific hardware support or conditional constraint in the scenario.