Anomalous Events Detection in Frequent Sequence Video via Object Segmentation Using Motion Pixel Compensation and Gradient

The surveillance takes major part in the present world. The collection of data through surveillance is not the thing required by the people but to use it effectively for decision support. Manual processing of huge amount of surveillance data is not feasible, a support to analyse and process all incoming data is needed by the operator of the surveillance system. In this proposal, an approach to intelligent video surveillance is introduced with significance on finding behavioural anomalies. The proposed technique of robust and efficient anomaly detection is capable of dealing with crowded scenes. Initially, the foreground segmentation of input frames is done to analyse the foreground objects and to effectively ignore irrelevant background dynamics. The input frames are split into non overlapping cells, and then the features are extracted based on motion, size and texture from each cell. Each and every feature type is independently analysed to detect the presence of an anomaly. An effective estimate of object motion is obtained by computing the optical flow of only the foreground pixels. To process even large training datasets, the motion and size feature are modelled by an approximated version of kernel density estimation.