1. Introduction
Detecting anomalies in surveillance videos, e.g. stalled objects, accidents and abnormal objects, has been a challenging task because of the shortage of annotated or labeled data and the variable video scenes. Therefore, it is almost infeasible to acquire the orbits of every object in the videos and then judge whether the orbit should be classified as anomaly or not. However, it is a truly valuable task due to its potential application in real world.