Multi-Dimensional Optical Flow Embedded Genetic Programming For Anomaly Detection In Crowded Scenes

Anomaly detection is an important issue in the fields of video behavior analysis and computer vision. Genetic Programming (GP) performs well in those applications. Histogram of Oriented Optical Flow (HOOF), which is based on Optical Flow (OF), is a significant method to extract features of frames in videos and has been widely used in computer vision. However, OF may produce a large number of features that will lead to high computational cost and poor performance. Moreover, HOOF accumulates optical flow values for moving objects in a region, so motion information of these objects may not be well represented. Especially in crowding scenes, common anomaly detection methods usually have a poor performance. Aiming to address the above issues, we propose a new feature called Multi-Dimensional Optical Flow (MDOF) and a new method GP-MDOF which embeds HOOF features and optimizes the structure of GP, and makes better use of the change information between consecutive frames. In this paper, we apply GP-MDOF to classify frames into abnormality or not. Experimental evaluations are conducted on the public dataset UMN, UCSD Ped1 and Ped2. Our experimental results indicate that the proposed feature extraction method MDOF and the new method GP-MDOF can outperform the popular techniques such as OF and Social Force Model in anomaly detection in crowded scenes.