Fast and reliable identification of abnormal crowd behaviour in surveillance footage

Abstract

Detecting unusual crowd events in surveillance video is crucial due to increased crime rates in recent years. However, automatically identifying these events is challenging because feature comparison across the training and test phases is time-consuming. To address this, we introduce the Markov-nearest transition based unusual events classifier (MTUEC) classifier, which classifies input frames as either normal or unusual events. The MTUEC algorithm comprises several modules: the spatial slice model, static object removal computation, spatio-temporal estimation, and the Markov-nearest transition based unusual events classifier. The MTUEC algorithm focuses on comparing features between immediate training frames and the testing frame. If the immediate training frame does not match the testing frame, the algorithm considers other training frames for comparison. This approach significantly reduces the time needed to detect unusual crowd events. To examine the effectiveness of the proposed method, we used two benchmark datasets for unusual crowd events: UMN and UCSD Ped1 and Ped2. We also compared the performance of the proposed approach with several existing algorithms for unusual event detection.