A hybrid unsupervised-weakly supervised method for video anomaly detection

A hybrid unsupervised-weakly supervised method for video anomaly detection

연구 분야: Artificial Intelligence

학회: Pattern Analysis and Applications

Video anomaly detection has recently become a critical field of research because of its potential applications in autonomous surveillance systems. We propose Hybrid-ConvATFM, a novel method that uniquely operates in both unsupervised and weakly supervised settings, making several key contributions. First, it alleviates the human annotation problem by employing a GMM-based divisive clustering algorithm to generate video-level pseudo-labels, enabling fully unsupervised operation when frame-level or video-level annotations are unavailable. Second, Hybrid-ConvATFM excels at capturing the complex temporal dynamics of anomalous events through its innovative architecture that combines convolutional autoencoders with attention modules, extracting enhanced local and global temporal dependencies of video segments. Third, our approach incorporates a temporal feature magnitude learning technique that effectively identifies and classifies abnormal segments with high precision. The versatility and effectiveness of Hybrid-ConvATFM is demonstrated through extensive experimentation on benchmark datasets. In the weakly supervised setting, our method achieves area under the curve scores of 97.70% on ShanghaiTech, 84.91% on UCF-Crime, and 83.51% on XD-Violence. In the fully unsupervised setting, Hybrid-ConvATFM attains 89.50% on ShanghaiTech, 79.61% on UCF-Crime, and 78.43% on XD-Violence, establishing new performance benchmarks in video anomaly detection across multiple operational scenarios.