CroSA: Unsupervised domain adaptation abnormal behavior detection via cross-space alignment

Huiyu Mu, Xianyu Zuo, Jiashuai Su, Yun Hu, Shubing Han, Baojun Qiao, Cheng Yuan, Lanxue Dang

Abstract: Most advanced video anomaly detection (VAD) methods have achieved excellent performance on public benchmark datasets from the same source. However, most VAD methods are sensitive to distribution-out data (i.e. data distribution is different), due to domain gaps caused by various interference factors (e.g. scene changes, camera parameters). To solve the domain gaps problem, previous methods mainly explored domain alignment, which requires collecting a large number of training samples. Instead, we proposed a Cross-Space Alignment (CroSA) unsupervised domain adaptation anomaly detection method, which utilized an auxiliary source domain dataset with the same data space as the target domain but different distribution as a guide to transfer the learned knowledge to the target domain. First, to effectively mine the behavioral correlation features related to the anomaly detection task in the case of large domain style differences, a feature decoupling method based on the domain invariant feature and domain specific style feature decomposition module is proposed, by simulating diverse normal distributions. Then, a multi-dimensional information cooperative alignment strategy is constructed, for cross-space domain alignment of pixel-level and feature-level. A multi-domain adversarial discriminative method is proposed to obtain multi-domain invariant feature representations. Finally, the target domain data pseudo-labels are iteratively generated by self-training, to further improve the detection effect of the CroSA model. Experimental results on several public benchmarks were shown balanced classification performance for VAD.