Efficient crowd anomaly detection using C3D-LSTM networks with enhanced attention mechanisms

Abstract

The rising deployment of surveillance systems in urban environments necessitates efficient automated anomaly detection methods. While showing promise, current deep learning approaches struggle with computational complexity and real-time performance in processing spatiotemporal information. This paper presents a hybrid framework integrating Convolutional 3D Networks (C3D), Long Short-Term Memory (LSTM) networks, and attention mechanisms for anomaly detection. Through a systematic evaluation of four attention mechanisms—self-attention, multi-head attention, Bahdanau attention, and Luong attention—we demonstrate their operational differences and their differential impact on feature extraction and classification performance across three diverse benchmark datasets. Our multi-head attention variant achieves state-of-the-art results with 99.40 % accuracy and 99.96 % Area Under the Curve (AUC) on Violent Flows, while maintaining robust performance across varying dataset complexities, achieving 91.87 % accuracy on the ShanghaiTech Campus and 79.7 % accuracy on the UCF-Crime dataset. Comprehensive cross-dataset evaluation demonstrates consistent improvements of 2.4 %–3.5 % over baseline approaches, with all attention mechanisms outperforming traditional spatiotemporal models. The proposed architecture effectively balances computational requirements with detection performance, maintaining real-time processing capabilities suitable for operational deployment. This framework advances the technical capabilities of anomaly detection systems while providing a validated foundation for practical deployment in diverse surveillance environments, from controlled scenarios to challenging real-world conditions.