Motion instability based unsupervised online abnormal behaviors detection

Automatically detecting anomaly in surveillance videos is a crucial issue for social security. Motion instability based online abnormal behaviors detection has been developed in an unsupervised way. The motion instability is composed of direction randomness and motion intensity of particles gotten by optical flow based consecutive motion feature extraction. The direction randomness is gotten based on weighted average of a circular variance of all particles. The motion intensity is determined according to average energy of all particles considering the camera perspective effect. A feature tracking based scheme has been employed to extract spatial-temporal motion features from videos to increase the processing speed. An adaptive dynamic thresholding strategy is developed to detect deviation of the track from the patterns observed both in direction randomness and motion intensity. Besides a double-threshold inference strategy is adopted to determine the range of the motion instability. A state transition model is used to reduce false alarm for confirming anomaly. The anomaly in the video is fast online detected in an ordinary hardware from a cluttered scene without any hypothesis for the scenario contents in advance. Comparative study with state-of-the-arts has indicated the superior performance of the developed approach.