Multiple object tracking (MOT) in dynamic environments poses significant challenges due to occlusions, varying object appearances, and complex motion patterns. These difficulties are further compounded in surveillance videos with low quality and distant targets, especially when abnormal movements and events occur within the scene. To address these challenges, this paper proposes two novel frameworks that integrates scene information as prior knowledge into the MOT process. Using scene information enhances object association and tracking accuracy. The first algorithm, called SceneAware Adaptive Tracking Algorithm (SAATA), improves MOT by introducing a targetcentric search area and employing the novel state assessment techniques for comprehensive scene analysis. The state assessment aspect involves a thorough analysis of the targets’ current state to a more accurate and context-aware tracking system that can handle occlusion. One notable advantage of SAATA is its ability to significantly reduce identity switches and track fragmentation compared to the conventional algorithms. The second algorithm, called Abnormality-Aware Tracking Algorithm (AATA), leverages the scene\\\\\\\\\\\\\\\'s normality or abnormality to label detections and tracks, incorporating this insight as prior knowledge into the construction of the cost matrix for association. Consequently, if a target exhibits abnormal behavior, the tracker is Journal Pre-proof alerted and adjusts its actions to prevent loss of the abnormal target. AATAdemonstrates enhanced performance in terms of multiple object tracking accuracy (MOTA), compared to other trackers, showing its ability to effectively track abnormal targets. Furthermore, through experiments and evaluations on diverse datasets, the robustness and efficiency of the two algorithms are demonstrated.

Keywords