A lightweight algorithm for video abnormal behavior detection by fusing wavelet transform and depthwise separable convolution
Jie Kang, Jiale Wang, Ting Xu, Mo Wang · 2025
To address the challenges of high computational cost and parameter redundancy of video anomaly detection models in resource-constrained embedded systems, this paper proposes a lightweight algorithm that integrates wavelet transform and depth-separable convolution. Based on the GAFN baseline using an encoder-decoder architecture, we first reconstruct the standard convolution using a 3D depthwise separable convolutional attention module at encoder layers 3-4 and decoder layers 1-2. This decomposition method enhances the selectivity of spatial channel features. In addition, we introduce a 3D wavelet convolution at encoder layer 4, which utilizes the multi-scale nature of the wavelet transform to extend the receptive field and improve fine-grained feature extraction. Experiments demonstrate that our model achieves 95.0% and 83.5% AUC on UCSD Ped2 and CUHK Avenue datasets, respectively, and the number of covariates is only 1.0M, which is kept at 16.7% of the benchmark, indicating that the algorithm maintains high detection accuracy for video anomalous behaviors while the model is lightweight.