Online Action Detection in Surveillance Scenarios: A Comprehensive Review and Comparative Study of State-of-the-Art Multi-Object Tracking Methods

Online action detection in surveillance scenarios presents considerable challenges, particularly due to the dynamically changing environments and real-time processing requirements. Within this context, Multi-Object Tracking (MOT) serves as a critical component of the online action detection pipeline. Despite the emergence of several state-of-the-art (SOTA) object trackers in recent years, a gap remains in the comprehensive evaluation of these trackers specifically for action detection in surveillance scenarios. This paper bridges this gap by offering a thorough study of SOTA MOT trackers, aimed at determining the influential factors affecting their performance in surveillance settings and identifying the trackers optimally suited for an online action detection pipeline. For relevance and rigor, we introduce SurvTrack, a new dataset derived from a subset of VIRAT-dataset explicitly designed for action detection tasks-but intended for object tracking. SurvTrack is utilized to assess these trackers under various conditions, including differing image resolutions and detector confidence thresholds. This study uncovers the distinctive strengths and weaknesses of each tracker, providing invaluable insights for researchers and practitioners in surveillance and action detection. Importantly, this work focuses on tracking methods within the action detection domain, underscoring the development of a tracker explicitly designed for action detection on pertinent datasets, such as VIRAT.