Abnormal object representation based on surprise model

The method of saliency map model in visual attention model is only applied to static images, lacking temporal measurements, and not suitable for video sequence. This paper presents a novel method of abnormal object representation based on surprise model in video. Spatial-Temporal Interest Points are extracted as candidate points for detection firstly, and motion features of candidate points, such as motion magnitude and direction, are obtained by optical flow approach. We exploit a computational method combining both spatial surprise and temporal surprise, which measures the spatial-temporal variation degree between prior knowledge and posterior knowledge. If the variation exceeds beyond a predefined threshold, the candidate point is discriminated as an anomaly. Experimental results show that our algorithm is robust, practical and implemented easily.