Attention-free temporal modelling for efficient video violence detection

Abstract Video-based violence detection is a critical component of intelligent surveillance systems, particularly in smart environments where real-time responsiveness and computational efficiency are paramount. Conventional deep learning models often rely on attention mechanisms to capture temporal dependencies across video frames, but these approaches introduce significant computational overhead, limiting their deployment on resource-constrained devices. This paper investigates whether the multi-head attention component of an established spatiotemporal violence-detection architecture can be replaced with a lightweight 1D temporal convolution while retaining competitive classification performance and reducing computational complexity. The proposed architecture retains the 3D convolutional DenseNet-style backbone and downstream temporal and classification components, while substituting the attention-based temporal modelling stage with a compact Conv1D module. The experimental results on four benchmark datasets demonstrate superior performance where the proposed model achieves 100% accuracy on Peliculas, 95.00% on Hockey, 95.50% on Violent Flow, and 93.50% on RLVD. Additionally, the proposed model reduces the total trainable parameter count by approximately 23% compared with the attention-based reference model, while the Conv1D temporal module itself requires approximately 2.67× fewer parameters than the multi-head attention module. The proposed model also converges faster during training and requires less inference time, making it ideal for real-time applications. The proposed model offers a scalable solution suitable for various applications like smart cities, transportation hubs, and public venues where rapid threat detection is essential. The proposed model also represents a significant step towards attention-free, efficient video violence detection for resource-limited environments.