Probabilistic Detection of Crowd Events on Riemannian Manifolds

Event detection in crowded scenarios becomes complex due to articulated human movements, occlusions and complexities involved in tracking individual humans. In this work, we focus on crowd event (activity) detection and classification. We focus on active crowd (continuously moving crowd) events. First, event primitives such as motion, action, activity and behaviour are defined. Furthermore, a distinction is made among event detection, action recognition and abnormal event detection. Further, event detection and classification are defined on Riemannian Manifolds that yields six different probabilities of the event occurring. Using a new probabilistic approach, an automated event detection algorithm is proposed that temporally segments the event using a novel framework. The results indicate that the proposed approach delivers superior performance in selected cases and similar results in other cases, whilst the detection model delay allows operation in near real-time. The Performance Evaluation of Tracking and Surveillance (PETS) 2009 dataset was used for evaluation. Existing crowd event detection approaches used supervised approach, whereas we eschew semi-supervised approach.