Activity analysis based on SOM

This paper presents a novel SOM algorithm to classify the moving objects activities according to their trajectories. Firstly, a trajectory is represented as a sequence of vector that consists of the temporal motion feature and predictive motion information of moving object, and a good classification result benefits from the predictive motion information. Secondly, the motion features and predictive information of normal trajectories are learnt by a SOM network, and a SOM network is constructed to pattern the similarity of normal moving trajectories. Finally, this SOM network is used to classify the normal or abnormal trajectories of the moving objects by detecting abnormal points of trajectories, especially at the exact moment once the abnormal activity occurs. Experiments show that the proposed algorithm is effective.