Spatial neighborhood-enhanced framework for efficient loss and regularization in skeleton-based anomaly detection

Skeleton-based video anomaly detection has emerged as a privacy-preserving alternative to appearance-based surveillance, yet existing methods face a fundamental trade-off: they either converge slowly in structured environments or overfit rapidly in complex ones. We present Spatial Neighborhood-Enhanced Loss (SNE), a lightweight auxiliary objective that resolves this dilemma by exploiting a simple spatial prior - nearby individuals tend to move similarly. SNE adapts its role to scene characteristics: in structured datasets (Avenue, STC), it accelerates convergence by 20.8-86.2% while improving accuracy (+0.85-3.30% AUC); in complex datasets (UBnormal), it acts as a regularizer that prevents catastrophic overfitting through a novel three-phase learning dynamic. Cross-domain experiments confirm SNE's robustness, maintaining 92.9% performance retention and achieving a +2.43% average improvement across diverse deployment scenarios.