Trajectory learning for event recognition in video surveillance
L. Ng, Hong Siang Chua · Swinburne figshare (Swinburne University of Technology) · 2012
This paper presents an automated event classification and anomaly detection system for vision based surveillance system. Events are observed as a sequence of angle and distance vectors extracted from the motion trajectories and a compact feature vector representation of event is proposed. An unsupervised approach in event learning is introduced with the aids of the well-established Hidden Markov Model (HMM) framework. The spatial-temporal dynamics of the events are modeled by using a distinct HMM for each event. Contextual information of the motion trajectory extracted from point of interest is also incorporated in this work to improve the accuracy of the event recognition system. Extensive empirical investigation for the proposed method showed the robustness of the system in event classification and anomalous detection.