Weakly Supervised EfficientNet B7 based Framework for Detecting Anomalies in Surveillance Scenes
Rahul Chiranjeevi. V, D. Malathi · 2025
In Recent days surveillance cameras are generating vast amount of data, which necessitates in developing an automated model for detecting abnormalities and anomalies in surveillance videos. In this Research a novel method is proposed using EfficientNet B7 for detecting anomalies like pedestrian fights, fires, accidents and throwing of objects. Proposed method processes the video data by converting the videos in frames, resizing and frame normalization. These frames are fed to EfficientNet model for detecting the features. A temporal feature aggregation mechanism is utilized to further refine the detection of anomalies by analysing the sequential frame dependencies. This approach balances the computational efficiency with high detection accuracy. Experiments were performed on various benchmark datasets like Chuk Avenue, Peds2 using different evaluation metrics like AUC, Precision, Recall and F1 Score. Results shows that a precision of 0.91, Recall of 0.93 and F1 score 0f 0.97 is achieved on the given datasets. This method diminishes dependence on big labelled datasets, rendering it scalable for large-scale surveillance implementations. Despite ongoing problems including fluctuating illumination conditions and privacy issues, this study represents a notable progression in deep learning-based anomaly detection for surveillance security.