SSGT: Spatio-Sequential Graph Transformer for Skeleton-Based Person Re-Identification and Out-of-Set Detection

Abstract

Reliable person re-identification for robotics and surveillance requires appearance-independent methods capable of rejecting unseen individuals. We propose a unified framework for 3D skeleton-based person ReID with integrated open-set detection. Unlike prior methods that model spatial joint relations and temporal dynamics separately, the proposed Spatio-Sequential Graph Transformer (SSGT) encodes intra-frame skeletal topology and inter-frame motion features and aligns them via graph correspondence using a dual-graph attention architecture. Topology-aware positional encoding and multi-head attention yield compact, discriminative motion embeddings robust to viewpoint changes and occlusions. To enhance identity separability and structural consistency, we introduce a dynamically balanced multi-objective learning scheme combining a prototype-based contrastive loss with masked spatio-temporal reconstruction. The adaptive weighting stabilizes training and balances competing objectives. Open-set recognition is performed in the learned embedding space, enabling reliable rejection of unknown individuals without retraining. Experiments on two public benchmarks and a new multi-view dataset show consistent improvements over state-of-the-art methods in open-set settings and robustness under short temporal windows suitable for real-time deployment.