A Survey on 3D Skeleton Based Person Re-Identification: Taxonomy, Advances, Challenges, and Interdisciplinary Prospects

1 Introduction

Person re-identification (re-ID) is an essential pattern recognition task of matching and retrieving a person-of-interest across different views or scenes, which has been widely applied to security authentication, smart surveillance, activity monitoring, healthcare, and embodied AI [Nambiar et al., 2019; Ye et al., 2021; Bao et al., 2026]. Recent economical and precise skeleton-tracking devices (e.g., Kinect [Shotton et al., 2011]) have simplified the acquisition of 3D skeleton data, enabling them to be a prevalent and versatile data modality for gait analysis and person re-ID [Liao et al., 2020; Rao et al., 2024]. Unlike conventional person re-ID methods that rely on appearance or facial characteristics [Ye et al., 2021], 3D Skeleton based person Re-ID (SRID) models typically exploit body-structure features and motion patterns (e.g., gait [Murray et al., 1964]) from 3D positions of key body joints to identify different persons. With unique merits such as small input data, light-weight models, privacypreserving without using appearances, and robustness against view and background variations [Han et al., 2017], SRID has attracted surging attention from both academia and industry [Rao et al., 2022].

1 Corresponding Author

In recent years, research on SRID has gained significant momentum, leading to diverse skeleton modeling and learning paradigms. Early endeavors [Barbosa et al., 2012; Munaro et al., 2014a; Andersson and Araujo, 2015; Pala et al., 2019] mainly extract hand-crafted features such as skeleton descriptors in terms of anthropometric, geometric and gait attributes of body. As these methods often require domain expertise such as anatomy and kinematics [Yoo et al., 2002] for skeleton modeling, they lack the ability to fully mine latent high-level features beyond human cognition. To resolve this challenge, recent mainstreammethods [Liao et al., 2020; Huynh-The et al., 2020; Rao et al., 2022; Rashmi and Guddeti, 2022] leverage deep neural networks to automatically perform skeleton representation learning for SRID. One of exemplar methods (termed "sequence-based modeling") is to model sequential dynamics and motion semantics from raw or normalized skeletons (e.g., joint trajectory) based on long short-termmemory (LSTM) and its variants [Wei et al., 2020; Rao et al., 2022]. However, they rarely investigate the intrinsic body relationships such as inter-joint motion correlations, thereby possibly overlooking some valuable skeleton patterns. Another paradigm (termed "graph-based modeling") mitigates this challenge by constructing skeleton graphs to model discriminative structural and actional features based on the interrelations of body parts [Rao and Miao, 2023]. This often requires multi-granularity body modeling and efficient relational reasoning mechanisms (e.g., collaborative learning) based on skeleton graphs. Despite the great progress of SRID, this rapidly evolving technique still lacks a systematic review, making it difficult for researchers to gain a holistic view of this field and embark on new research endeavors.

In light of this, we present the first survey on SRID, elucidating recent advancements of skeleton modeling, learning paradigms, evaluation benchmarks, current challenges, and interdisciplinary applications. Firstly, we define the SRID task and provide a milestone overview to illustrate the origin and key advancements of SRID as shown in Fig. 2a. Secondly, we propose a systematic taxonomy of SRID methods to categorize them into hand-crafted, sequence-based, and graph-based modeling, and elaborate on their foundational mechanisms and representative approaches. We also illustrate the basic definitions and common methods within three mainstream SRID paradigms (supervised, self-supervised, unsupervised). Thirdly, we introduce existing public benchmarks, evaluation metrics, and protocols for SRID, while comprehensively evaluating state-of-the-art methods across different benchmarks to compare their performance and efficiency. Meanwhile, we conduct qualitative analysis of different methods to compare their key properties with a discussion of advantages and disadvantages. Finally, we discuss the current challenges in SRID and identify potential directions for future research. An overview of promising SRID applications in interdisciplinary areas, spanning healthcare, embodied AI, and security, is further provided and illustrated with a case study. The structure of this survey, including skeleton modeling (Sec. 3), learning paradigms (Sec. 4), benchmarks and evaluation (Sec. 5), challenges and prospects (Sec. 6) is shown in Fig. 3. We hope our survey can bring new insights to researchers and expedite future research in SRID.

2 Overview

2.1 Task Description

As illustrated in Fig. 1, the input of SRID task is a 3D skeleton sequence that belongs to a certain pedestrian, and the output is the predicted identity. Formally, we denote a 3D skeleton sequence as S = (s1, · · · , sf) ∈ Rf×j×3 , where st ∈ Rj×3 denotes the tth skeleton with 3D coordinates of j body joints. Each skeleton sequence S corresponds to a person identity I, where I ∈ {1, · · · , C} and C is the number of different classes (i.e., identities). In the SRID task, we generally have training set, probe set, and gallery set, respectively denoted as ΦT = {STi }N1 i=1 , ΦP = {SPi }N2 i=1 , and ΦG = {SGi }N3 i=1 that contain N1, N2, and N3 skeleton sequences of different persons collected from different scenes or views. The task target is to learn a hand-crafted, sequencebased or graph-based model (detailed in Sec. 3) that maps 3D skeleton sequences into effective representations, so that we can query the correct identity of an encoded skeleton sequence representation in the probe set via matching it with the sequence representations in the database (i.e., gallery set). SRID is essentially a retrieving and matching problem.

2.2 Origin and Advancements

As presented in Fig. 2a, the first SRID research [Munaro et al., 2014a] using hand-crafted skeleton descriptors commenced in 2014, coming after the first RGB video based method [Farenzena et al., 2010] and before the first depthbased approach [Haque et al., 2016]. Then, Rao et al. proposed the first deep learning based SRID paradigm in 2020, followed by the first self-supervised, unsupervised, and supervised paradigms [Rao et al., 2020; Rao and Miao, 2022a; Rao and Miao, 2023]. Over the past five years, an increasing number of innovative models have been devised specifically for SRID and related emerging tasks, including LSTM

models (e.g., AGE [Rao et al., 2020]), CNN models (e.g., SkeletonGait [Fan et al., 2024]), Transformer models (e.g., TranSG [Rao and Miao, 2023]), GCN models (e.g., CTRGCN [Chen et al., 2021]), MLP models (e.g., SimMC [Rao and Miao, 2022a]), and hybrid/composite models (e.g., SMSGE [Rao et al., 2021a]) (compared in Table 2).

2.3 Taxonomy of SRIDMethods

As shown in Fig. 3, we propose a systematic taxonomy for SRID approaches based on skeleton modeling (see Sec. 3) and learning paradigms (see Sec. 4). For skeleton modeling, we divide them into three categories, including (1) hand-crafted modeling using manually-extracted features (e.g., skeleton descriptors), (2) sequence-based modeling that focuses on sequential features (e.g., joint trajectory) of 3D skeletons, and (3) graph-based modeling that represents 3D skeletons as graphs, and further subcategorize them by different learning focuses such as body relations or graph dynamics. In terms of learning paradigms, we group them into (1) supervised SRID paradigms that require skeleton labels for feature learning, (2) self-supervised SRID paradigms that combine pretext tasks for skeleton representation learning with labeled fine-tuning, and (3) unsupervised SRID paradigms that learn skeleton features without using labels.

3 SkeletonModeling

We elaborate on different skeleton modeling including their foundational mechanisms and representative approaches.

3.1 Hand-CraftedModeling

Gait Attributes. Extracting gait features is a common way to characterize unique walking patterns of an individual [Cunado et al., 2003], typically including (1) kinematic parameters (e.g., angles of hips, knees, and feet), and (2) spatiotemporal parameters (e.g., stride length, gait cycle time, velocity). They can be manually computed with:

where xi, yi, zi denote 3D coordinates of the ith joint, the set Ψ defines adjacent joints constrained by the human skeleton model, and two kinds of angles αij and βij are calculated from these joint pairs. The velocity can be calculated by

where

We calculate the step length by averaging the highest values of the difference between the right and left feet, and adopt the mean stride length over all n strides following [Andersson and Araujo, 2015].

Anthropometric Attributes. The Euclidean distance between two joints, such as bone lengths, limb dimensions, height, are usually computed as the anthropometric features (fA) to differentiate individuals [Barbosa et al., 2012]:

where Ji=(xi, yi, zi), Jj=(xj, yj, zj), Ψ∗ not only contains adjacent joints but also can be extended to cover more anthropometric properties, such as 13 (D13) or 16 (D16) skeleton descriptors in [Munaro et al., 2014b; Pala et al., 2019].

Geometric Attributes. The incorporation of geometric skeletal measurements, including body-part ratios and interjoint geodesic distances on the mesh surface, can enhance feature representations in person re-ID [Barbosa et al., 2012].

The above hand-crafted features are often learned by different classifiers (e.g., KNN) to perform person re-ID [Andersson and Araujo, 2015; Nambiar et al., 2017]. They are also combined with different metric algorithms [Pala et al., 2015] or other modalities such as 3D point clouds [Gharghabi et al., 2015; Bondi et al., 2016; Pala et al., 2019; Munaro et al., 2014a] to further boost person re-ID accuracy.

3.2 Sequence-BasedModeling

Joint Motion. The walking patterns are characterized by the motion of joints, which can be defined as the differences between body joint positions [Liao et al., 2020]:

where st = {J t1 , J t2 , · · · , J t j }, J t i = (xti , yt i , zt i ) denotes the 3D coordinates of ith joint in the tth skeleton, and i ∈ {1, 2, · · · , j}. The occluded or masked joints during walking can also help models learn useful motion patterns [Rashmi and Guddeti, 2022].

Pose Dynamics. The consecutive skeletons typically conveys dynamics of unique body poses, which can be encoded sequentially by temporal learning models (e.g., LSTM):

where ϕ(·) denotes the model function to encode the pose dynamics of skeletons, ht-1 represents the latent representation of previous (t-1) poses, which provides the temporal context information to encode the long-term pose dynamics ht for person re-ID [Rao et al., 2022]. The joint distances, relative joint positions, and bones are combined in [Wei et al., 2020; Zhang et al., 2023] to enhance pose dynamics learning.

Semantics Learning. By encoding the skeleton sequence into latent high-dimensional representation, latent motion semantics such as motion continuity can be captured to enhance feature learning. Representative semantics learning tasks include skeleton sequence reconstruction and prediction, which can be simply formulated as follows:

The reverse reconstruction (Recon.) and prediction require the model function ϕ(·) to capture high-level semantics (e.g., order and correlations) to achieve the target output, which facilitates learning more meaningful gait representations for person re-ID [Rao et al., 2022]. A few studies also explore the semantics learning of motion consistency [Rao and Miao, 2022a] and cross-scale body relations [Rao et al., 2021a].

Importance Inference. Different skeletons and their feature representations typically possess different importance in characterizing poses and discriminative patterns of a person, which can be exploited to mine key skeletons or hard samples [Hermans et al., 2017] for SRID learning. To aggregate key skeleton features for SRID, AGE [Rao et al., 2020] explores a locality-aware attention mechanism to integrate features of important skeletons in the sequence, while SM-SGE [Rao et al., 2021a] infers key correlations between different body-joint nodes. Rao et al. [2024] further propose a hard skeleton mining mechanism to adaptively infer importance of multi-level skeleton representations for SRID.

3.3 Graph-BasedModeling

Body Relations. Skeletons can be naturally modeled as graphs based on the physical connections of human body joints. For each skeleton xt, we represent it as the graph G t (V t , E t ), where V t = {vt1 , vt2 , · · · , vt j }, vt i ∈ R3 , i ∈ {1, · · · , j} and edges E t = {eti,j | vt i , vt j ∈ V t }, eti,j ∈ R. E t denotes the set of connections between different joints. Based on the skeleton graphs, we can model the inherent correlations, such as limb collaboration [Rao et al., 2021b; Rao and Miao, 2022b] and channel-specific relations [Chen et al., 2021] for SRID. The graph motifs can also be devised to enhance the structure and gait related relation learning [Rao and Miao, 2025].

Graph Dynamics. The dynamics of skeleton graphs can be leveraged to learn and capture the temporal evolution of joints' connection patterns and limbs' coordination for SRID. The process can be simplified as

where fG(·) denotes the graph encoding model (e.g., graph convolutional networks [Fu et al., 2023]), ϕ(·) is the temporal learning model (e.g., LSTM) to encode the long-term dynamics of graph representations. To facilitate spatio-temporal graph learning, self-supervised pretext tasks, such as graph reconstruction and prediction, are incorporated to learn more effective skeleton features [Rao and Miao, 2023].

Multi-Level Graphs. Multi-level skeleton graph representations are devised to characterize coarse-to-fine body structure and motion [Li et al., 2020], and various graph learning tasks such as sparse graph prediction and cross-scale graph inference [Rao et al., 2021a] are proposed to help learn different-level graph semantics. Based on graph transformers, masked reconstruction [He et al., 2022] and motifs [Rao and Miao, 2025] are further explored for high-level structural (e.g., locality) and gait semantics learning.

4 Learning Paradigms

Supervised SRID paradigms leverage skeletal annotations or labels (e.g., identity class) to guide the model to learn discriminative features, typically using cross-entropy (CE) loss as follows:

where Ii,j denotes the ground-truth identity label (Ii,j = 1 iff the ith sample belongs to the j th identity otherwise 0), ˆIi,j is the probability that the ith sample is predicted as the j th identity, and N1 is the number of training samples. This loss has been widely applied to classic supervised models such as SVM [Munaro et al., 2014b], KNN [Munaro et al., 2014a; Gharghabi et al., 2015; Nambiar et al., 2017], MLP [Andersson and Araujo, 2015], Adaboost [Pala et al., 2019], random forest [Elaoud et al., 2021] to learn skeleton descriptors to classify different individuals. Recent studies explore supervised skeleton prototype learning paradigms by employing the contrastive loss

where Sk,j denotes the j th skeleton sequence belonging to the kth identity, pk represents the prototype generated by the feature centroid of k-class skeletons, and τ is the temperature for contrastive learning. Based on the losses in Eq. (10) and (11), diverse network architectures such as skeletonbased LSTM [Wei et al., 2020; Rashmi and Guddeti, 2022], CNN [Liao et al., 2020; Huynh-The et al., 2020], and Transformer [Rao and Miao, 2023] are trained for SRID.

Self-Supervised SRID paradigms usually combine unlabeled learning of pretext tasks (e.g., skeleton semantics learning) and labeled fine-tuning. A general form of pretext objective can be formulated as

where T1(·), T2(·) denote the sequence transformation functions (e.g., reverse and shift in Eq. (8)) to set the input and target of a pretext task, ϕ(·) represent the model encoding function, and Dis(·) is the distance metric (e.g., Euclidean distance). By employing sequence-based pretext tasks such as sparse sequential prediction [Rao et al., 2021b], or graphbased tasks such as structure-trajectory prompted reconstruction [Rao and Miao, 2023], self-supervised learning encourages the model to capture high-level motion concepts and class-related spatio-temporal semantics for SRID.

Unsupervised SRID paradigms perform skeleton representation learning without using any labels. Existing methods mainly adopt unsupervised skeleton prototype learning frameworks, which use the feature centroids of clustering as prototypes (see Eq. (11)) [Rao and Miao, 2022a; Rao et al., 2024]. Their performance could be affected by the robustness of clustering algorithms and contrastive learning mechanisms, while unlabeled pretext tasks are often combined to enhance general skeleton semantics learning.

5 Benchmarks and Evaluation

Benchmark Datasets. Table 1 summarizes the data source, identity number, skeleton amount, and viewpoint number of commonly-used datasets for SRID evaluation. They can be mainly categorized to two types: (1) Sensor-based datasets, where 3D skeleton data are captured from depth sensors such as Kinect, and (2) RGB-estimated datasets, in which skeleton data are estimated from RGB videos using 2D or 3D pose estimation models [Chen and Ramanan, 2017; Cao et al., 2019]. Existing SRID datasets typically contain skeleton data collected from varying scenarios such as multiple views (KS20, KinectREID), appearance and clothing changes (BIWI RGBD-ID, IAS-Lab RGBD-ID), and different illumination conditions (KGBD), which enables a comprehensive evaluation of both short-term and long-term SRID performance. In addition to standard SRID datasets, the 3DGait dataset provides an interdisciplinary benchmark for evaluating the generalizability of SRID models on healthcarerelated tasks, such as neurodegenerative disease prediction (see Sec. 6.2) [Rao et al., 2025b].

Evaluation Metrics and Protocols. In SRID, the performance is typically evaluated based on several mertics, including Cumulative Matching Characteristics (CMC), Rank-1 accuracy (R1), Rank-5 accuracy, Rank-10 accuracy, and Mean Average Precision (mAP) [Zheng et al., 2015]. Multiple evaluation protocols are employed across varying datasets, including probe-gallery matching evaluation (main protocol evaluated in Table 2), RGB-estimated evaluation (representative protocol evaluated in Table 3), random view evaluation, cross-view evaluation [Rao et al., 2024], zero-shot crossdataset evaluation [Rao et al., 2022], etc.

Comparison of Performance and Efficiency. As shown in Table 2, D16 and PM are two most competitive handcrafted methods, performing well on both BIWI and IAS benchmarks. However, recent deep learning based models significantly surpass them: The latest sequence-based model (Hi-MPC) and graph-based model (MoCos) achieve superior performance to them across different datasets. Notably, the top three performers-GPGait, TranSG, and MoCos-all utilize skeleton graph representations, highlighting the efficacy of graph modeling and body relation learning for SRID task. In terms of efficiency (cf. Fig. 2b), the sequence-based SimMC demonstrates the most compact design requiring only 0.15M parameters, followed by the graph-based MG-SCR (0.32M). Generally, methods utilizing Transformer architectures (GaitTR, TranSG, MoCos) exhibit significantly smaller parameter sizes than using other architectures such as LSTM. It is noteworthy that CNN-based approaches such as PoseGait incur substantially higher computational cost.

Comparison of Key Characteristics. As summarized in Table 2, hand-crafted methods design features based on domain knowledge, and employ well-established machine learning models, offering relatively high explainability and low model complexity. By contrast, sequence-based methods learn directly from raw data to capture sequential dynamics, but often overlook valuable features within body structure and joint relations. Graph-based approaches are designed to model these structural and motion connections via pre-defined topologies. By quantifying joint importance and collaboration, graph models provide a more interpretable framework. While both sequence-based and graph-based paradigms can integrate pretext tasks (e.g., reconstruction) and multi-scale modeling (e.g., multi-level body representations) to boost learning, only 4 out of 14 (28.6%) of these methods provide theoretical analyses of their effectiveness.

6 Challenges and Prospects

In this section, we elucidate key challenges in SRID data and models with a discussion of potential directions. We also present promising interdisciplinary applications of SRID.

Data Scarcity and Imbalance. Existing 3D skeletons are mainly collected from prevailing depth sensors such as Kinect [Shotton et al., 2011], while diverse skeleton collection settings (e.g., different devices in uncontrollable environments) have not be thoroughly explored. In contrast to existing RGBbased person re-ID data (e.g., MSMT17 [Wei et al., 2018]), the available skeleton data for person re-ID are relatively scarce and imbalanced. As shown in Table 1, existing Kinectbased SRID datasets contain 4.8 to 205.8 thousand skeletons with less than 200 identities, while the numbers of skeletons belonging to each identity often differ greatly (i.e., imbalanced class distribution). Training on these datasets might incur limited generalization ability of data-driven SRID models using deep neural networks, with a risk of model over-fitting.

Data Noise. Kinect-based and RGB-estimated skeleton data possibly contain noise due to: (1) Internal factors: The key points are generated by device built-in algorithms (via depth images) or pose estimation models (via RGB images), thus the inherent limitation of algorithms (e.g., precision, robustness) might result in inaccurate skeletons; (2) External factors: The quality of skeleton data may also be affected by device's tracking distance, illumination changes (e.g., influence structured light in Kinect V1), and source data quality (e.g., image resolution). Such inherent noise puts high demand on the model robustness against random perturbations.

To address these two challenges, higher-quality 3D skeleton data should be collected or generated. It is feasible to devise skeleton denoising models and augmentation strategies for skeleton generation. GAN-based pose generators [Yan et al., 2017] and diffusion models [Ho et al., 2020] could be transferred to generate and denoise 3D skeleton data. The future efforts include (a) collecting and opening new larger-scale SRID datasets, (b) transferring existing skeleton datasets from other areas to person re-ID datasets and formulating appropriate evaluation protocols, (c) estimating 2D/3D skeletons from large-scale public person re-ID datasets to construct new estimated SRID datasets, so as to advance SRID research and related person re-ID community.

Model Robustness. A few SRID models report unstable performance variations that are sensitive to different model parameter initialization, hyper-parameter settings (e.g., clustering parameters) and varying data distributions [Rao and Miao, 2022a]. As these studies opt for the best-trained model with finely-tuned parameters, this practice often fails to reflect the architecture's true average performance and shows limited adaptability in real-world applications.

A key future direction is to systematically investigate how key factors (e.g., model initialization, data quality) affect robustness, which is essential for developing more reliable models. Theoretical analyses of performance variations in terms of model-approximated functions and convergence conditions can also be provided for more robust model design. Moreover, multi-faceted evaluation metrics, such as performance average and standard deviation, should be reported to better measure the overall robustness of models.

Model Generalizability. Most SRID models are trained on a single dataset with limited data sizes, views, scenes or conditions. As a result, they can hardly generalize to different data domains (e.g., vulnerable to domain shifts) or real-world scenarios (e.g., RGB-estimated skeleton data).

A potential solution is to exploit larger-scale SRID datasets to train the model across diverse scenarios, so as to learn more domain-general skeleton semantics. It is also feasible to explore domain adaptation or generalization techniques to cotrain and transfer models. More benchmarks should be investigated for generality evaluation of SRID models.

Interpretability. Current SRID models typically lack intuitive explanations for the effectiveness of model architectures, skeleton features, and prediction results. This opacity not only poses risks of erroneous outcomes but also hinders their reliable applications at scale.

To this end, different human-friendly explanation including pose/feature visualization and corresponding languagebased description can be considered. There also exist various architecture-specific explanation mechanisms, e.g., class activation maps (CAM) [Zhou et al., 2016] for CNN, knowledge graphs [Ji et al., 2021] for graph neural networks, which could be applied to explainable skeleton learning. Moreover, large language models (LLMs) can serve as an agent to cotrain the SRID model and interpret the learning process. A possible direction is to transform skeletons into time series text or pose images, and leverage LLMs or other agents to analyze the pose/feature importance.

6.1 Emerging and Open Directions

Multi-Modal Learning. Combining 3D skeletons and emerging modalities such as images [Bao et al., 2025], radio frequency waves (e.g., mmWave) is a promising direction, as they can provide pose or gait information from different dimensions (e.g., silhouettes) to better perform recognition tasks. To this end, it is crucial to develop unified gait representations that generalize effectively across different data modalities.

Cross-Modality Evaluation Protocol. To enable a fair comparison of person re-ID methods on multi-modal (e.g., RGB-D) benchmarks, it is imperative to formulate a crossmodality evaluation protocol that standardizes re-ID settings (e.g., probe/gallery settings, single/multi-shot recognition) of skeleton-based, depth-based, radar-based and other methods.

Skeleton Foundation Model. LLMs and pose generative models [Lucas et al., 2022; Bao et al., 2024] can be synergized to build a skeleton foundation model by training it on large-scale skeletons. It can hopefully be fine-tuned for diverse tasks such as skeleton visualization, augmentation, prediction, etc. Such foundation model will help investigate the limit of adaptability, generality, and interpretability of 3D skeletons, and can potentially serve as a healthcare copilot [Zhao et al., 2025b] with expert rules and evidence-based reasoning [Zhao et al., 2025a; Zhao et al., 2021] to empower proactive health monitoring.

Privacy Protection. A unique advantage of SRID models is that they inherently protect privacy by avoiding the use of explicit appearance data. However, the illegal or irresponsible deployment of person re-ID technologies could jeopardize public security, making it crucial to establish SRID-related laws to protect social safety and personal privacy.

6.2 Interdisciplinary Prospects

As shown in Fig. 4, we list several promising applications of SRID in terms of healthcare, embodied AI, and security:

Healthcare. Since physiological and psychological states are intrinsically correlated with walking patterns (e.g., Parkinsonian gait) [Lu et al., 2025; Rao et al., 2025a], the pre-trained SRID models can be transferred for neurodegenerative disease diagnosis and affective/psychological detection (e.g., depression) [Lu et al., 2023a; Lu et al., 2023b]. Moreover, they can hopefully support medical gait analysis and rehabilitation assessment (e.g., stroke) via non-intrusive monitoring in daily environments.

To demonstrate the above interdisciplinary applicability, we follow [Rao et al., 2025b] to systematically evaluate various SRID models on a representative healthcare task: Neurodegenerative disease prediction. The case study in Table 4 shows that these models can capture generalizable discriminative gait patterns for this task, suggesting their promising potential for broader real-world healthcare applications.

Embodied AI. With smaller input data and lower resource requirement than RGB-based models, SRID can serve as a fundamental semantic link for intelligent agents and their interaction systems: Agents can act via skeleton-based avatar motion retargeting, perceive through human-centric intention prediction, interact using skeletal gesture based control, and evolve within skeleton-driven synthetic environments, which facilitate robust identity-aware collaboration in robotics.

Security. Leveraging the robustness of skeletal features against appearance variations [Han et al., 2017], SRID advances public safety through privacy-preserving authentication and cross-modal (e.g., face, fingerprint) identity verification. It also enables forensic behavioral analysis to track suspects and detect abnormal events across varying scenarios particularly where visual details are lacking.

Sports Science. By capturing fine-grained skeletal dynamics, SRID enables biomechanical analysis for athletic performance optimization (e.g., running gait correction and swimming stroke refinement [Zhou et al., 2026]) and sports injury prevention through abnormal movement pattern detection. Furthermore, it provides real-time skill acquisition feedback for interactive coaching (e.g., pose comparison with expert templates), thereby supporting data-driven training in both competitive and recreational sports.

7 Conclusion

In this paper, we provide the first comprehensive review of 3D skeleton based person re-ID (SRID). We first define the SRID task and present a timeline to overview the major advancements of SRID. Then we formulate a taxonomy of SRID approaches in terms of different skeleton modeling and learning paradigms, systematically presenting foundational mechanisms and reviewing representative approaches. An empirical evaluation of state-of-the-art SRID methods is conducted across various benchmarks and evaluation protocols to compare their effectiveness, efficiency, and key properties. We further discuss critical challenges along with potential directions, and highlight interdisciplinary prospects of SRID.

Acknowledgments

This research is supported by the RIE2025 Industry Alignment Fund - Industry Collaboration Projects (IAF-ICP) (Award I2301E0026), administered by A*STAR, as well as supported by Alibaba Group and NTU Singapore through Alibaba-NTU Global e-Sustainability CorpLab (ANGEL). This research is also partially supported by the National Research Foundation, Singapore under its AI Singapore Programme (AISG Award No: AISG2-PhD/2022-01-034[T]).