An investigation of videos for abnormal behavior detection

Abstract

The identification of abnormal behavior has several applications. There are several ways, ranging from classical to deep learning based. It may be used to monitor campuses, banks, transportation, and airports. In many circumstances, the context determines whether real-life events are common or unusual. Recent video surveillance anomaly detection systems are good enough, but they come at a significant computational cost and require particular hardware resources. When it comes to real-time anomaly detection, extra emphasis must be paid on lowering model complexity, which causes computational and memory demands. This study attempts to find real-world abnormalities in CCTV recordings, such as violence, detention, property destruction, assault, burglary, blast and theft. The impact of these abnormalities on public safety is enormous. The study also provides a low-cost algorithm for detecting crowd irregularities. The proposed model uses the convolutional neural network based on DenseNet121 as the feature extractor. Our suggested framework has an AUC of 86.63 percent on the UCF-Crime dataset and obtains new stateof-the-art performance