Video Anomaly Detection in Crime Analysis Using DenseNet-121

Contemporary crime analysis significantly depends on video anomaly detection (VAD) for immediate surveillance and swift threat identification. This research presents a deep learning methodology employing DenseNet-121 for precise and effective anomaly detection in surveillance video. The strongly linked convolutional layers of DenseNet-121 provide comprehensive spatial-temporal feature extraction, enhancing anomaly classification precision. The model is trained and assessed utilizing a preprocessed UCF Crime dataset comprising several anomalous acts, including theft, assault, and vandalism. The suggested method incorporates a temporal attention mechanism alongside motion-based preprocessing to improve the identification of small anomalies in intricate environments. Experimental results indicate enhanced performance, attaining 92.8% accuracy, 91.5% precision, 90.2% recall, and an F1-score of 90.8%, surpassing traditional CNN-based models. The results underscore the model's efficacy in practical criminal detection contexts, enhancing intelligent surveillance systems. This research enhances automated video analysis for public safety and situational awareness by aiding law enforcement agencies in crime prevention and security improvement.