Intracluster Contraction for Pedestrian Anomaly Detection Under One-Class Classification Setting

Xiaolei Wang, Xiaoyang Wang, Huihui Bai, Eng Gee Lim, Jimin Xiao

Abstract: Video anomaly detection (VAD) focuses on identifying unusual or unexpected events within video sequences. However, fully supervised learning approaches are ill-suited for this task because of the scarcity of anomaly examples. To tackle this challenge, one-class classification (OCC) has emerged as an effective solution for VAD. Under the OCC framework, the training phase requires only normal samples to build a detection model. Therefore, the core of VAD in the OCC setting lies in effectively modeling the normality of pedestrian behaviors during training. Recent studies often rely on appearance features for accurate feature modeling. Nevertheless, an over reliance on appearance-only features makes the detection system insensitive to pedestrian movement anomalies. In this study, we combine general scene-appearance features with a simple velocity representation derived from optical flow to achieve high-performance VAD. We posit that velocity information, when combined with a well-designed distribution modeling, can effectively describe pedestrian activities. Given the diversity of normal pedestrian motion, we employ a multimodal probabilistic model to learn the distribution of velocity normality during training. Directly fitting on the original velocity features, however, leads to mixed clusters and dispersed cluster boundaries. This impedes the separation between normal and abnormal samples. To address this, we propose an intracluster contraction (ICC) module. The ICC maps the extracted features to the neighborhoods of their local high-density centers, forcing more compact local clusters. This significantly enhances the distinction between normal and abnormal samples. For deep scene-appearance features, we also apply the ICC to adjust the feature distribution and design a distance-based density estimator to detect appearance anomalies. Supported by both appearance and motion information, our method demonstrates high adaptability to complex and variable scenarios. It achieves competitive results on three benchmark datasets.

External IDs:dblp:journals/tai/WangWBLX26