Improving Robbery Behaviour Potential Assessment in Indoor Security Cameras through Optimized Loitering Detection with Custom YOLOV5-DeepSORT

Abstract Predicting Robbery Behavior Potential (RBP) in indoor surveillance is crucial for proactive security. Building upon the foundational multi-modal RBP prediction framework by Pouyan et al. [Original Paper Ref], which utilizes head cover, crowd, and loitering detection fused by a fuzzy inference system, this paper addresses a key limitation in its loitering analysis component, particularly for low-resolution footage. We introduce an optimized loitering detection module by integrating a custom-trained YOLOV5 model for robust person detection with the DeepSORT tracking algorithm. This enhancement aims to improve tracking accuracy in challenging low-resolution scenarios, leading to more reliable loitering cues. The refined loitering information, alongside the original head cover and crowd metrics, is then processed by the fuzzy inference engine to yield a more accurate RBP assessment. Preliminary evaluations (or Expected evaluations) on the UCF-Crime dataset suggest that this targeted improvement in the loitering module can significantly enhance the overall F1-score for RBP prediction and subsequent robbery detection tasks.