Construction and Benchmarking of a Dataset for Human Pose Estimation of Suspicious Behavior in Outdoor Environments

Automated detection of suspicious human behavior in outdoor public spaces is a critical challenge for intelligent surveillance systems. Appearance-based approaches raise privacy concerns and degrade under occlusion and lighting variation. Skeleton-based Human Pose Estimation (HPE) offers a privacy-preserving alternative by characterizing behavior through body joint trajectories. However, large-scale annotated datasets targeting suspicious pose sequences in real outdoor environments are scarce. This paper presents SUSP-POSE, a novel benchmark dataset containing 8,400 pose-annotated sequences of normal and suspicious behaviors across six outdoor public location types. A formal taxonomy of 12 suspicious behavior categories derived from criminological literature and expert consultation is defined. Baseline experiments using ST-GCN, CTR-GCN, and PoseFormer establish benchmark results: binary detection F1 = 0.820 (CTR-GCN), 13-class macro F1 = 0.710. The dataset, annotation tools, and evaluation code are released publicly to foster reproducible, privacy-aware behavioral analysis research.