P2-VGAN: a lightweight velocity-weighted graph attention network for privacy-conscious edge anomaly detection

Introduction Two persistent challenges hinder intelligent video surveillance in public spaces: dense pixel-based models incur heavy computational costs on edge infrastructure, and continuous capture of identifiable imagery creates substantial visual privacy liabilities. Skeleton-based pipelines, which abstract each video frame into articulated landmark coordinates, support visual data minimization by discarding appearance information. However, conventional graph backbones weight joint trajectories uniformly across time, frequently confusing benign high-speed locomotion (such as sprinting for transit) with violent anomalies. Methods This paper presents the Privacy-Conscious Velocity-Weighted Graph Attention Network (P2-VGAN; the prefix “P2” is retained as a legacy identifier designating privacy-conscious representation rather than a cryptographic guarantee), an efficient framework engineered for desktop-class CPU edge gateways. The pipeline decouples motion analysis from visual appearance by extracting 33 planar body landmarks via a single-shot pose estimator; the application does not intentionally store or transmit decoded RGB frames after pose extraction, ensuring downstream processing operates strictly on appearance-reduced pose coordinates. The primary contribution is the Velocity-Weighted Attention Mechanism (VW-AM), an ultra-lightweight gating module requiring only two shared parameters that converts inter-frame coordinate displacements into a bounded squared-displacement motion proxy. This proxy modulates node features prior to multi-head graph attention, directing message passing toward rapidly moving limbs while damping near-stationary joints. Trained under a weakly supervised Multiple Instance Learning objective with temporal smoothness and sparsity regularizers, P2-VGAN is evaluated across five random seeds on the ShanghaiTech Campus and a human-centric subset of UCF-Crime. Results On ShanghaiTech, the model achieves a frame-level AUC of 84.72 ± 0.28% with 1.81 M downstream trainable parameters (pose-extractor parameters excluded) and sustains 34.01 frames per second (FPS) on an Intel Core i7-10700 CPU without hardware acceleration. When compared directly with an identical Vanilla GAT backbone without gating (83.14 ± 0.39%), VW-AM provides an unconfounded +1.58% absolute AUC gain ( p = 0.00096). On the human-centric UCF-Crime subset (215 test videos), it attains 77.88 ± 0.35% micro-average AUC, with high-motion categories (Fighting, Assault) exceeding 82% AUC. Stepwise ablations document an incremental +2.52% gain from VW-AM over an adaptive ST-GCN baseline. Discussion We position P2-VGAN as a Pareto-efficient framework for lightweight CPU surveillance, acknowledging that coordinate streams reduce visual surface exposure but do not eliminate biometric gait re-identification.