Multiscale squeezed normalizing flow network for video abnormal behavior detection
Tianyu Wei, Liancheng Su, Weili Ding · 2025
Real-time detection of human anomalous behaviors in video surveillance poses significant challenges. To address the limitations of conventional normalizing flow networks in multi-scale spatio-temporal feature extraction, local detail preservation, and noise sensitivity, we propose a Multi-scale Squeezed Normalizing Flow (MSS-NF) framework based on pose estimation. The network incorporates a multi-scale dilated fusion mechanism that employs parallel dilated convolutions to extract hierarchical features across channel-spatial dimensions. This design expands the receptive field while maintaining spatial resolution, thereby enhancing multi-scale feature representation capabilities. Integrated with a squeeze-and-excitation module, the framework dynamically recalibrates feature weights to amplify critical patterns and suppress noise interference, effectively mitigating overfitting risks. Experimental evaluations demonstrate that MSS-NF achieves state-of-the-art detection accuracy and robustness on both ShanghaiTech and UBnormal datasets, exhibiting particularly superior performance in fine-grained anomaly recognition compared to mainstream approaches.