Detecting the unseen: Micro-movement and social context integration for suspicious behavior detection

Video anomaly detection is a key research area in intelligent surveillance. Existing methods are mainly effective for detecting short-term, visually salient anomalies such as fighting or running, but they often struggle to identify prolonged, subtle, and intentionally concealed behaviors. In practice, offenders frequently loiter and observe target areas over extended periods as part of pre-crime reconnaissance. Detecting such suspicious behaviors early enables proactive crime prevention. To address this challenge, we propose a novel task: Suspicious Behavior Detection (SBD). The core difficulty lies in identifying long-term, intention-driven anomalies that lack explicit visual cues. Inspired by psychological and criminological theories, we design a local–global perception framework. The local module uses ST-GCN and Bi-LSTM to extract skeletal micro-movements and gaze patterns, while the global module integrates long-term trajectories with social-semantic features such as regional concealment and visibility. A bottleneck attention mechanism fuses multi-scale features for unsupervised detection. To support this task, we also construct a dedicated dataset, SBDataset. Experiments on both public datasets and SBDataset show that our method provides consistent improvements over mainstream approaches. In particular, the framework achieves notable gains on SBDataset, highlighting its ability to capture concealed intent-driven behaviors, while maintaining competitive performance on challenging public benchmarks. These results demonstrate the promise of incorporating social semantics into anomaly detection and the potential of SBD as a new research direction.