ABNORMAL BEHAVIOR DETECTION IN SURVEILLANCE SYSTEMS USING RESNET50: A TRANSFER LEARNING APPROACH

Abstract

With the proliferation of cameras and surveillance systems in metropolitan areas, a notable concern for the scientific community in computer vision revolves around devising automated techniques to interpret video-based scenes. Activities of particular interest may indicate potential security threats to people, places, or objects. Most techniques related to this subject have concentrated on recognizing people’s actions and activities. Consequently, these methods identify any abnormal actions that might be considered suspicious. This study develops an automated technique for detecting abnormal behavior in surveillance systems and explores the effectiveness of transfer learning using the ResNet50 architecture. This study improves the accuracy and efficiency of abnormal behavior detection in video-based scenes, thereby enhancing the ability of surveillance systems to identify potential security threats. This model is fine-tuned using three distinct datasets (UCSD (Ped1, Ped2) and CUHK Avenue), comprising both normal and abnormal behaviors. Experimental results show that our transfer learning approach achieves state-of-the-art performance on all three datasets, outperforming several baseline methods. Our approach demonstrated good generalization ability, performing well on a separate test set from each surveillance dataset. The novelty of this study lies in the proposed transfer learning approach for abnormal behavior detection in surveillance systems. This study demonstrates state-of-the-art performance on multiple datasets by leveraging a pre-trained ResNet50 model and fine-tuning with diverse datasets, including normal and abnormal behaviors.