Person re-identification via 3D skeletons is an emerging research area that is attracting increasing attention within the pattern recognition community, according to a comprehensive survey published on arXiv. Authored by Haocong Rao and Chunyan Miao, the paper provides a systematic taxonomy of 3D skeleton based person re-identification (SRID) methods, reviews state-of-the-art advances, and highlights challenges and interdisciplinary prospects. For enterprise technology leaders, SRID offers a privacy-preserving and illumination-invariant biometric alternative to traditional camera-based identification, with potential applications in secured facility access, retail analytics, and workforce tracking within supply chain environments.
Taxonomy of SRID Methods
The survey defines the SRID task and categorizes existing methods into three main modeling paradigms: hand-crafted, sequence-based, and graph-based. Hand-crafted methods rely on manually designed features from skeleton joint coordinates. Sequence-based approaches leverage recurrent neural networks (RNNs) or transformers to model temporal dynamics of skeleton sequences. Graph-based methods represent the human skeleton as a graph, with joints as nodes and bones as edges, and apply graph convolutional networks (GCNs) to capture spatial structure and temporal evolution. The paper elaborates on each category with foundational mechanisms and representative models.
Learning Paradigms and Benchmarks
The survey covers three mainstream learning paradigms: supervised, self-supervised, and unsupervised. Supervised methods learn discriminative representations from labeled skeleton sequences. Self-supervised approaches design pretext tasks (e.g., skeleton reconstruction, contrastive learning) to learn general features without manual annotations. Unsupervised methods aim to identify individuals across camera views without any identity labels, often using clustering and iterative refinement. A thorough evaluation of state-of-the-art SRID methods is conducted over various types of benchmarks and protocols to compare their effectiveness, efficiency, and key properties.
Key Challenges and Prospects
Despite significant progress, the survey identifies several challenges: handling occlusions, variations in skeleton quality due to depth sensor noise, and generalizing across different datasets and environments. The authors also note the need for larger, more diverse datasets and more robust evaluation protocols. Looking forward, they highlight interdisciplinary applications of SRID with a case study, pointing to potential integration with biometrics, human-computer interaction, and security systems. For technology procurement leaders, these developments signal a maturation of skeleton-based biometrics that could soon offer reliable, low-latency person identification for high-security zones, although real-world deployment will require further validation on domain-specific datasets and integration with existing surveillance infrastructure.