Spatial–Temporal Decoupled Invertible Neural Networks for Skeleton-Based Video Anomaly Detection

Skeleton-based video anomaly detection has gained significant attention in security and forensics applications due to its privacy-preserving properties that abstract human motion without capturing identifiable biometric information. However, existing methods typically process spatial pose configurations and temporal motion dynamics in an entangled manner, limiting both interpretability and discriminative power. We observe that anomalies manifest distinctly in spatial and temporal dimensions: spatial anomalies correspond to unusual body configurations while temporal anomalies reflect irregular motion patterns. To explicitly model this decomposition, we propose Spatial-Temporal Decoupled Invertible Neural Networks (STD-INN) that separates spatial and temporal information within normalizing flow architectures for precise anomaly detection. Specifically, STD-INN introduces a learnable invertible decomposition mechanism that adaptively splits encoded skeleton features into spatial components capturing pose structures and temporal components encoding motion dynamics. We then model them using dual independent normalizing flows: the spatial flow incorporates graph convolutional networks to leverage skeletal topology, while the temporal flow employs causal temporal convolutions to capture sequential dependencies. A lightweight cross-calibration module enables controlled information exchange between the two streams. Extensive experiments on benchmark datasets demonstrate that STD-INN achieves state-of-the-art performance with lightweight parameters while offering improved diagnostic interpretability compared to existing approaches.