Enhancing Video Surveillance and Anomaly Detection with Deep Learning Solutions in Dynamic Environments

Abstract

This research fills the gaps of Convolutional Neural Networks (CNNs) for timely detection of dynamic abnormalities by presenting a novel video anomaly detection model integrating ConvGRU and CNNs and Gated Recurrent Units (GRUs). In order to offer efficient handling of spatial and temporal data in films, temporal dependencies are modeled by ConvGRU while spatial features are learned by ResNet50. By employing motion analysis and entropy filtering to detect focused frames, the proposed solution significantly reduces computation expenses. In preparing to generate anomaly probabilities for classification, the model structure initially extracts space features with ResNet50, followed by time features with ConvGRU. The model does a very good job, as indicated by its performance on the UCF Crime dataset, with 300 movies across five actions. Its accuracy on the validation set is 95.12%, its validation loss is 0.2103, and its Area Under the Curve (AUC) score is 0.9823. Moreover, at a cost of 2.10 × 10¹¹ FLOPs, the COD-3D ResNet model also beats other high-end models, such as P3D and Q3D, when it comes to classifying correctly as well as for computing. Especially in detecting gunfire (94% recall) and assault (99% accuracy), the model is very accurate and recalls well. Overall, the proposed hybrid architecture offers an extremely successful and computationally efficient real-time video anomaly detection system suitable for surveillance.