Skip to main content An official website of the United States government Here's how you know Here's how you know Official websites use .gov A .gov website belongs to an official government organization in the United States. Secure .gov websites use HTTPS A lock ( Lock Locked padlock icon ) or https:// means you've safely connected to the .gov website. Share sensitive information only on official, secure websites. Search Log in Dashboard Publications Account settings Log out Search… Search NCBI Primary site navigation Search Logged in as: Dashboard Publications Account settings Log in Search PMC Full-Text Archive Search in PMC Journal List User Guide PERMALINK Copy As a library, NLM provides access to scientific literature. Inclusion in an NLM database does not imply endorsement of, or agreement with, the contents by NLM or the National Institutes of Health. Learn more: PMC Disclaimer | PMC Copyright Notice BMC Sports Sci Med Rehabil . 2026 Mar 4;18:180. doi: 10.1186/s13102-026-01612-0 Search in PMC Search in PubMed View in NLM Catalog Add to search Assessment of basketball players’ motion quality degradation by video-based virtual sensing and the AGCN-Mamba network Zhi Yang Zhi Yang 1 School of Physical Education, the College of Arts and Sciences·Kunming, Kunming, 650222 China Find articles by Zhi Yang 1 , Yuanye Zhao Yuanye Zhao 2 School of Information Engineering, the College of Arts and Sciences·Kunming, Kunming, 650222 China Find articles by Yuanye Zhao 2, ✉ Author information Article notes Copyright and License information 1 School of Physical Education, the College of Arts and Sciences·Kunming, Kunming, 650222 China 2 School of Information Engineering, the College of Arts and Sciences·Kunming, Kunming, 650222 China ✉ Corresponding author. Received 2025 Dec 31; Accepted 2026 Feb 13; Collection date 2026. © The Author(s) 2026 Open Access This article is licensed under a Creative Commons Attribution-NonCommercial-NoDerivatives 4.0 International License, which permits any non-commercial use, sharing, distribution and reproduction in any medium or format, as long as you give appropriate credit to the original author(s) and the source, provide a link to the Creative Commons licence, and indicate if you modified the licensed material. You do not have permission under this licence to share adapted material derived from this article or parts of it. The images or other third party material in this article are included in the article’s Creative Commons licence, unless indicated otherwise in a credit line to the material. If material is not included in the article’s Creative Commons licence and your intended use is not permitted by statutory regulation or exceeds the permitted use, you will need to obtain permission directly from the copyright holder. To view a copy of this licence, visit http://creativecommons.org/licenses/by-nc-nd/4.0/ . PMC Copyright notice PMCID: PMC13067531 PMID: 41782160 Abstract Traditional physiological monitoring methods in basketball games are often invasive and struggle to capture long-term fluctuations in movement quality. This study presents a framework for assessing movement quality degradation based on video-based virtual sensing and an Adaptive Graph Convolution Network (AGCN)-Mamba network. First, the framework extracts human skeleton sequences via pose estimation and transforms them into dynamic feature tensors resembling Inertial Measurement Unit (IMU) data, thereby simulating the data stream of wearable devices. The AGCN module then captures coordination relationships between non-adjacent limbs, while Mamba’s selective scanning strategy enables long-term fatigue tracking throughout a game with low computational complexity. Experimental results under a subject-disjoint protocol demonstrated a Top-1 action recognition accuracy of 94.82%, and the movement quality scores correlated strongly with expert evaluations (correlation coefficient r = 0.91). A per-frame latency of 12.08 ms further indicates the framework’s applicability in non-invasive movement monitoring. These results show that combining adaptive graph learning with linear state-space modeling allows effective extraction of deep features and efficient edge computation. The proposed method provides a solid theoretical and technical foundation for the development of next-generation, low-power, and high-precision intelligent wearable fitness monitoring systems. Keywords: Fitness assessment model, Virtual wearable sensing, Adaptive graph convolutional network, Mamba state space model, Action recognition Introduction In modern high-intensity basketball, athletes’ performance exhibits dynamic fluctuations due to fatigue. Extended competitions or high-load training not only reduce technical stability and tactical execution, but also significantly increase the risk of non-contact sports injuries [ 1 , 2 ]. Traditional physical monitoring methods, such as post-game physiological measurements or subjective fatigue scales, are often delayed and influenced by individual bias, making real-time feedback and control difficult [ 3 ]. With the integration of computer vision and micro-sensing technologies, extracting kinematic degradation phenotypes from non-invasive video data has emerged as an important approach for evaluating athletes’ performance potential. This method captures fatigue-related behavioral changes by identifying “micro-compensation actions,” which refer to non-standard movement patterns that athletes adopt to maintain specific performance goals (e.g., shooting accuracy) under fatigue. For instance, increasing knee flexion to compensate for reduced upper-limb strength reflects adaptive adjustments in joint coordination. In biomechanics, such actions correspond to “adaptive adjustments of movement redundancy.” Although data-driven models have advanced action recognition, applying them to movement quality assessment still faces significant challenges. Existing skeleton-based models often focus on local feature extraction and struggle to capture spatiotemporal dynamics over long-duration games [ 4 ]. Moreover, variations among athletes and environmental factors—such as camera angle, occlusion, and lighting—introduce domain shift, which limits model generalization in real-world deployment. Current assessment frameworks also face a trade-off between maintaining high modeling accuracy and efficient long-sequence inference, and they lack consideration for real-time detection in edge computing scenarios [ 5 ]. To address these challenges, this study proposes a video-based virtual sensing AGCN-Mamba framework. The framework does not rely on physical sensors. Instead, it derives virtual acceleration and velocity signals from 2D video using computer vision algorithms. An adaptive graph convolution network (AGCN) dynamically captures coordination relationships between non-adjacent limbs, while a linear-complexity Mamba state-space model handles long-term motion sequences. This framework aims to provide a low-latency monitoring solution for movement quality degradation. Cross-subject experiments on the Basketball Action Recognition (BAR) and the NTU RGB + D dataset validate the robustness of this framework in extracting competitive behavioral features. The main contributions of this study are summarized as follows: ➢Video-based virtual sensing framework: A method to extract IMU-like features from 2D video was developed, enabling non-invasive quantification of kinematic indicators and avoiding interference from physical sensors on athlete performance. ➢AGCN-Mamba spatiotemporal modeling network: By combining adaptive graph learning with a linear state-space model, the framework effectively captures subtle movement quality degradation trends over long sequences while significantly reducing inference latency (12.08 ms). ➢Kinematic degradation assessment mechanism: A multi-task learning architecture integrates action classification with regression of “pseudo-ground truth” based on kinematic deviations, providing a reliable statistical proxy for performance monitoring in the absence of physiological feedback. ➢Comprehensive experimental validation: Extensive comparisons on a professional basketball dataset and the NTU RGB + D benchmark demonstrate the model’s generalization ability across subjects and its potential applicability in edge computing environments. Literature review Current status of basketball performance monitoring The evaluation and monitoring of basketball players’ physical fitness has long been a central topic in sports science. Traditional research primarily focused on the relationship between physiological function and competition performance, as well as on verifying the effects of training interventions. Petrigna et al. (2022) [ 6 ] highlighted the need for standardized testing protocols for specific movement patterns and pointed out that existing methods fell short in enabling real-time monitoring. Ibáñez et al. (2023) [ 7 ] examined the predictive value of physical fitness indices for professional female basketball players and found that aerobic capacity and explosive power were crucial for supporting high-level performance. Regarding training interventions, several studies systematically reviewed the positive effects of functional training on athletes’ physical characteristics and metabolic functions [ 8 – 10 ]. However, most research relies on intermittent sampling or post-match data, lacking technical means for real-time, non-invasive assessment of athletes’ competitive state. This limitation significantly constrains coaches’ ability to provide accurate tactical guidance during live competition. Review of deep learning-based action recognition and generalization Advances in the Internet of Things (IoT) and deep learning have opened new avenues for automatic action recognition using sensor and video streams, offering potential for real-time kinematic quality degradation assessment. Liu et al. (2025) [ 11 ] developed an IoT-based system capable of detecting basketball movements in real time, demonstrating the feasibility of on-site computation and analysis in sports environments. Considering the heterogeneity of basketball movements, Khobdeh et al. (2024) [ 12 ] combined the YOLO algorithm with a deep fuzzy long-short-term memory (LSTM) network to extract movement features and accurately distinguish different basketball actions. However, integrating fuzzy logic increased computational complexity. Zhang (2024) [ 13 ] proposed a fine-grained detection algorithm based on deep learning, while Wang et al. (2024) [ 14 ] applied 3D convolutional neural networks (3D CNNs) to capture spatiotemporal features, achieving promising results in motor skill classification. Nevertheless, domain shift remains a critical challenge in real-world scenarios. Tanveer et al. (2023) [ 15 ] emphasized that effective domain adaptation strategies were essential for improving model generalization across diverse users and complex environments. Current hybrid models still struggle to maintain stable evaluation performance over long sequences under highly variable competitive conditions. Current research on spatiotemporal modeling and edge computing efficiency To address the inherent limitations of traditional deep learning approaches, Graph Neural Networks (GNNs) and attention mechanisms have gradually become research hotspots. Lu (2024) [ 16 ] studied gesture recognition in multi-modal robots using a spatiotemporal graph convolution network (ST-GCN) and experimentally demonstrated that graph-structured data could effectively represent gesture samples. Building on this, Zahra et al. (2025) [ 17 ] proposed a dynamic GNN model for classifying group behaviors from a drone’s perspective, overcoming the constraints of static graphs in complex dynamic scenes. In the field of long-sequence modeling, Wang et al. (2024) [ 18 ] combined self-supervised learning with the Transformer architecture to analyze player trajectory data in depth, which improved the accuracy of offensive strategy recognition. However, such models still struggle to handle high-frequency, high-dimensional time-series data typical of basketball games. To resolve the trade-off between computational efficiency and long-sequence modeling accuracy, the Mamba architecture—based on linear-complexity State Space Models (SSMs)—offers clear advantages in motion recognition tasks. Wen et al. (2025) [ 19 ] integrated the Mamba module into a graph convolution network, enabling efficient fusion of spatiotemporal features and significantly reducing inference latency. Fazzari et al. (2025) [ 20 ] employed selective state-space modeling to suppress environmental interference and extract core action frames. At the same time, algorithmic efficiency on edge devices is critical. Inspired by Güney et al. (2022) [ 21 ], Güney and Bayilmis (2024) [ 22 ], Güney et al. (2022) [ 23 ], who implemented real-time action monitoring on mobile GPUs and edge devices, this line of research emphasized the importance of lightweight inference methods. Such methods needed to preserve long-sequence spatiotemporal modeling accuracy while remaining suitable for edge deployment. This objective constituted a central goal of the present study. Research gaps and advantages of the proposed framework The existing literature highlights several limitations in basketball physical fitness evaluation. Traditional fitness tests lack the capability for real-time and continuous monitoring. Moreover, widely used deep learning models, such as CNN-LSTM, struggle to capture the complex structure and long-term dynamics of the human musculoskeletal system. Although advanced Transformer-based models can achieve high accuracy, their substantial computational demands make them unsuitable for edge deployment. To address these challenges, this study proposes a novel framework that combines AGCN with Mamba. The framework is designed to provide a high-precision, real-time, and edge-deployable evaluation model for basketball players, effectively balancing modeling accuracy with computational efficiency and enabling practical on-site motion quality assessment. Research model Overall architecture of the fitness assessment model This study proposes an end-to-end deep learning framework designed to extract high-precision kinematic features from non-intrusive video data and quantify athletes’ fitness states. Considering the core application scenario—basketball action recognition in the BAR dataset, which includes diverse movements such as shooting and dribbling, as well as long-sequence characteristics—the model architecture places particular emphasis on capturing spatiotemporal features. To ensure that the backbone network generalizes well across different scenarios when processing human skeletal topologies, the framework was extensively validated on the large-scale international benchmark dataset NTU RGB + D. As illustrated in Fig. 1 , the overall workflow follows a hierarchical “perception–transformation–cognition” logic, designed to meet the stringent robustness requirements of complex competitive environments. Fig. 1. Open in a new tab Overall architecture of the fitness assessment model To overcome the limitations of a single video modality in physical quantification, the core task of the preprocessing module is to generate high-fidelity “virtual sensor” data streams. This technique functionally simulates the signal characteristics of real Inertial Measurement Units (IMUs) by computing skeletal point dynamics. In this study, the OpenPose algorithm was employed to accurately locate key points of the human body in each frame. The algorithm extracts the two-dimensional coordinates (x, y) and confidence score (C) for major body parts, including the head, neck, shoulders, elbows, wrists, hips, knees, and ankles. The detection results are then visualized for verification. To suppress the high-frequency noise inherent in monocular video, a Kalman filter was applied to the raw coordinate sequences prior to feature matrix construction. This smoothing step produced more stable proxy estimates of velocity and acceleration. The detailed processing workflow is illustrated in Fig. 2 . Fig. 2. Open in a new tab Virtual sensor nodes and topology based on human skeleton In order to make the model accurately identify the dependence of various parts of the human body, a human skeleton diagram G=(V , E) is constructed on the basis of the above virtual sensor nodes. Among them, the V set is the main joint node of the human motion chain in Table 1 , and the E set includes physical connections (knees connect ankles, and adaptive learnable edges are given in the subsequent AGCN algorithm, which is used to represent the non-physical dependence in the human body other than physical connections (such as wrists and knees when shooting; After feature engineering, each node contains a multi-dimensional feature vector ( x , y , c , V X , v y , a x , a y , … ) in each frame. These rich data dimensions provide a powerful data foundation for the model to identify the micro-deformation and compensation mechanism caused by fatigue. Table 1. Virtual sensor node definitions and feature channels Node ID Body Part (Region) Virtual Sensor Type Extracted Features (Channels) Physical Significance in Basketball 0–1 Nose / Neck Head IMU Position ( x , y ), Velocity ( v x , v y ), Acceleration ( a x , a y ) Head stability and core control during shooting 2–4 Right Shoulder / Elbow / Wrist Upper Limb IMU (Right) Shooting kinetic chain and arm muscle fatigue 5–7 Left Shoulder / Elbow / Wrist Upper Limb IMU (Left) Ball protection, balance, and defensive posture intensity 8–10 Right Hip / Knee / Ankle Lower Limb IMU (Right) Lower limb explosive output and jump height decay 11–13 Left Hip / Knee / Ankle Lower Limb IMU (Left) Landing stability and ACL injury risk 14–17 Left Eye / Right Eye / Left Ear / Right Ear Auxiliary Node Head orientation and visual focus tracking Open in a new tab Directly deriving dynamic features such as velocity and acceleration from 2D video keypoint coordinates often introduces high-frequency quantization noise. This process is also constrained by depth loss inherent in monocular vision. To mitigate the limitations of relying solely on visual input for physical quantification, a Kalman filter was applied to the raw coordinate sequences provided by OpenPose before constructing the feature matrix. This temporal smoothing effectively reduces spurious fluctuations in derivative estimates caused by occlusions or sudden detection jumps. Although virtual IMUs cannot fully recover absolute mechanical values in 3D space, the adaptive adjacency matrix in the AGCN automatically learns the relative motion patterns between body parts. Consequently, the model does not depend on the absolute accuracy of acceleration. Instead, it infers athletes’ physical fitness states by capturing changes in acceleration over time and coordinated oscillation patterns among joints. In this manner, the framework achieves a mathematically robust approximation of exercise load even in the absence of depth information. It should be emphasized that the proposed virtual sensing framework functions as a kinematic simulation rather than a direct replacement for physical sensors. Due to the perspective projection of monocular cameras, the method has inherent limitations in handling keypoint occlusions and depth ambiguity, and it cannot fully restore absolute biomechanical properties in three-dimensional space. Therefore, the model focuses on capturing relative temporal patterns and coordination variability among keypoints. Quantitative calibration of the virtual signals against real IMU data will be a key direction for future improvement. AGCN-based spatial modeling module After constructing the skeletal graph based on virtual sensors, an AGCN was used to extract topological features of human posture in each frame [ 24 , 25 ]. To avoid confusion with mathematical symbols, the original adjacency matrix defined by physical connections is denoted as . The core innovation of AGCN lies in extending the conventional fixed skeletal topology into a hybrid learnable model. According to Eq. ( 1 ), the output feature is determined jointly by the physical structure , a globally learnable matrix , and a data-driven dynamic matrix : 1 denotes the number of spatial partitions of the convolution kernel (typically root, centripetal, and centrifugal sets). is the weight matrix for the k -th partition, is the N×N adjacency matrix representing physical skeletal connections. is a fully parameterized learnable matrix initialized as zeros ( ) and optimized during training via backpropagation to reach the global optimal static topology , defined as (2): 2 This mechanism enables the model to identify latent motion patterns, such as the coordinated interaction between knees and wrists during a shot, effectively compensating for the limitations of modeling based solely on anatomical connections. To further capture unique, instantaneous inter-limb coordination variations in each frame (such as fatigue-induced micro-movement deformations), Eq. ( 1 ) incorporates a data-driven dynamic matrix . The spatial modeling module under the AGCN is illustrated in Fig. 3 . Fig. 3. Open in a new tab Spatial modeling workflow under the AGCN (B=Batch size, T=Time sequence, V=Joint nodes, C=Channels) The dynamic matrix is generated by computing the similarity between nodes using a Gaussian embedding function. Specifically, for the input feature , the model first maps it into a latent subspace through two independent embedding functions θ and ϕ , as expressed in Eqs. ( 3 ) and ( 4 ): 3 4 and represent the feature vectors of nodes i and j , respectively, while and are learnable linear projection parameters. Based on these projected features, the correlation between any two nodes is computed via a dot product and normalized using the Softmax function to obtain the elements of the dynamic adjacency matrix, as shown in Eq. ( 5 ): 5 Equation ( 5 ) ensures that the sum of each row of is 1, thus constructing a probabilistic attention map. Finally, to maintain gradient stability in deep networks and prevent degradation, the AGCN module integrates a residual connection architecture. The graph-convolved spatial features are added to the input features and passed through an activation function to produce the final spatial modeling output , as shown in Eqs. ( 6 ) and ( 7 ): 6 7 denotes batch normalization, and is the nonlinear activation function. Temporal modeling module analysis supported by Mamba To address the quadratic computational bottleneck in processing long-duration motion sequences, this study incorporates the Mamba SSMs [ 26 ]. By introducing a selective scan mechanism, Mamba overcomes the limitations of traditional s SSMs in handling complex content, achieving a dynamic balance between inference efficiency and long-range dependency modeling. Furthermore, Liu et al. (2025) [ 27 ] highlighted that Mamba and its variants, such as Vision Mamba, exhibit superior linear scalability in spatiotemporal modeling tasks compared with Transformer architectures, owing to a more efficient recursive mechanism. This characteristic enables the framework to capture subtle degradation trends in motion quality over extended sequences in the BAR dataset, while imposing minimal hardware overhead. Mamba models continuous kinematic processes using linear ordinary differential equations, mapping the input sequence to the output through a latent state . Its continuous-time formulation is expressed in Eq. ( 8 ): 8 is the evolution matrix, which governs the compression and memory of the system state; and are the input and output projection matrices, respectively. Compared with the Transformer model based on ordinary differential equations, it is consistent with the dynamic evolution law of nature. When reasoning in this way, the calculation can be completed efficiently by recursive means, and the memory consumption demand will be significantly reduced. In order to realize the mapping between the continuous-time model and the discrete sampled video skeleton data, it is necessary to discretize the system according to the Zero-Order Hold, ZOH) principle, introduce the time scale parameter , and transform the original continuous matrix into the corresponding discrete matrices and by using the parameters A and B, so as to realize the matching. This discretization allows the model to adapt to different sensor sampling rates, as shown in Eq. ( 9 ): 9 In the traditional SSM model, such as LTI system, the parameters of matrix , and are usually set to fixed and consistent values in time, which makes it difficult for this kind of model to change according to the time series characteristics of input signals. In contrast, the innovation brought by Mamba is that it adds the mechanism of selective scanning. In this way, the original fixed model parameters can be transformed into functional expressions of input variables, which has stronger adaptability. To maintain linear complexity when processing long-duration motion sequences, this study employs the Mamba SSMs. Before entering the Mamba block, spatiotemporal features are organized into a 3D tensor , where B denotes the batch size, T the sequence length, and D the feature dimension. The selective scan mechanism endows the model with content-aware capabilities through an input-dependent parameterization process. As shown in Eq. ( 10 ), the time-scale parameter , the input matrix , and the output matrix are dynamically generated from the current input : 10 Using these dynamic parameters, the model performs selective state updates. When corresponds to low-intensity background actions, such as standing during a dead ball, the small encourages the system to preserve historical states. Conversely, during high-intensity actions, such as jumps or sprints, a larger forces state updates, enabling the capture of critical kinematic degradation features. The discrete recursive update is expressed as Eq. ( 11 ): 11 Through parallel scan algorithms, Eq. ( 11 ) can be efficiently computed on GPUs, achieving several times faster training and inference than Transformers when handling long sequences in the BAR dataset. After temporal modeling with the Mamba layer, the output sequence features capture both action temporal dependencies and latent trends of physical fatigue through the evolution of hidden states. The data then enters the fusion and prediction layer. Here, global average pooling aggregates temporal features into a fixed-dimension context vector z, which is subsequently mapped to the task space via fully connected layers. To jointly perform action recognition and fitness assessment, a multi-task loss is designed, combining classification loss (cross-entropy) and regression loss (mean squared error, MSE) with weighting, as shown in Eq. ( 12 ): 12 is the action category probability output via Softmax, denotes the normalized fitness (or kinematic quality degradation) score, obtained through either a Sigmoid or linear activation function, and , are hyperparameters balancing the two tasks. This design ensures the model not only understands what action the athlete is performing but also quantifies the physical cost underlying the movement. The Mamba temporal modeling and prediction workflow is illustrated in Fig. 4 . Fig. 4. Open in a new tab Pseudocode workflow of mamba temporal modeling and prediction (Note: the SSM subscripts for matrices A, B, and C have been omitted for clarity.) Experimental results and discussion Experimental setup To comprehensively evaluate the effectiveness of the AGCN-Mamba model in extracting kinematic features and assessing fitness degradation, a multi-level experimental validation protocol was designed. Datasets The primary evaluation dataset is the publicly available BAR dataset ( https://github.com/hkair/Basketball-Action-Recognition ). It contains 2,800 high-definition video clips captured from multiple viewpoints, covering six core basketball actions: Shooting, Dribbling, Passing, Layup, Defensive Footwork, and Off-ball Running. The dataset features complex backgrounds, frequent occlusions, and significant viewpoint variations, closely simulating the conditions of high-intensity competition. To verify the backbone network’s generalization across different skeletal topologies, the study also employed the NTU RGB + D 60 dataset ( http://github.com/shahroudy/NTURGB-D ), a leading benchmark in skeleton-based action recognition. The official Cross-Subject (X-Sub) protocol was followed to test the model’s robustness in non-basketball scenarios using large-scale human pose data. Detailed dataset statistics are summarized in Table 2 . Table 2. Dataset Statistics for Experiments Feature Dimension BAR Dataset NTU RGB + D 60 Benchmark Subjects 15 professional athletes 40 volunteers Classes 6 core basketball actions 60 general daily actions Total Samples 2,800 video clips 56,880 skeleton sequences Avg. Length 120 frames (~ 4 s) 80 frames (~ 2.7 s) Frame Rate 30 FPS 30 FPS Purpose Core application and fitness evaluation Backbone network generalization validation Open in a new tab Semi-supervised generation of kinematic quality labels Since the original datasets lack quantitative fatigue annotations, a semi-supervised label generation strategy based on kinematic heuristics was designed. It should be noted that these labels represent a Kinematic Quality Attenuation Score, rather than direct physiological fatigue indicators. The procedure is as follows: three national-level basketball coaches were invited to select standard exemplar clips for each action category, defining the “peak performance” reference (assigned a score of 1.0). To assess the reliability of these expert-prior-based labels, statistical consistency tests were performed on the peak-performance clips and their scores, as shown in Table 3 . The inter-rater agreement coefficient (Kendall’s W) was 0.89, and the intra-class correlation (ICC) was 0.87, indicating high consistency among evaluators and establishing an objective foundation for the pseudo-ground truth generation. Table 3. Consistency analysis of coach-defined baselines Metric Value 95% CI p -value Interpretation Kendall’s W 0.89 [0.84, 0.93] < 0.001 Very High ICC 0.87 [0.81, 0.91] < 0.001 Excellent Open in a new tab Subsequently, Dynamic Time Warping (DTW) was applied to quantify deviations between the skeleton trajectories of test samples and the reference trajectories. These deviations were combined with variations in action duration and the vertical stability of the Center of Mass. Finally, the resulting kinematic deviation metrics were min-max normalized to the [0, 1] range. Lower scores indicate greater degradation in movement quality due to fatigue. This pseudo-ground truth, grounded in expert priors and kinematic indicators, effectively captures the external manifestation of fatigue in athletic performance. Experimental and hyperparameter settings To rigorously address reviewers’ concerns regarding potential data leakage, a strict subject-disjoint split was applied to the BAR dataset. Among the 15 athletes, 10 were allocated to the training set, while the remaining 5 were split between validation and test sets in a ratio of approximately 7:1.5:1.5. This ensures that the test set contains athletes—and their unique movement styles—that were never seen during training, guaranteeing the true generalization of evaluation results. Experiments were conducted in a high-performance computing environment, detailed in Table 4 . During preprocessing, OpenPose keypoint coordinates were smoothed using a Kalman filter. Data augmentation strategies such as random node dropping and Gaussian noise injection were applied to simulate extreme scenarios, including missing sensor signals or sudden visual detection jumps. Table 4. Experimental environment configuration Category Item Specification Hardware GPU NVIDIA GeForce RTX 3090 (24GB VRAM) CPU Intel Xeon Gold 6226R @ 2.90 GHz Software Deep Learning Framework PyTorch 2.1.0 / CUDA 12.1 Keypoint Extraction Engine OpenPose v1.7.0 (BODY25 model) Loss Function Weights (λ) Open in a new tab Baseline models and detailed configuration To fully evaluate the performance boundaries of the AGCN-Mamba model, several representative baseline models were selected, including CNN-LSTM, ST-GCN, standard Transformer, ST-GCN++ (Enhanced Spatial-Temporal Graph Convolutional Networks), and 2s-AGCN (Two-stream Adaptive Graph Convolutional Networks). These advanced baselines were chosen to provide a rigorous comparative validation, highlighting the advantages of the proposed adaptive graph modeling combined with linear state-space scanning architecture in handling complex, long-duration motion sequences. To maintain academic rigor, all baseline experiments were executed under identical hardware, preprocessing pipeline, and subject-disjoint split protocols. Hyperparameters for each baseline model strictly followed those recommended in the original papers to ensure fair comparison, without artificially limiting their performance. Detailed hyperparameter configurations are summarized in Table 5 . Table 5. Hyperparameter configuration comparison between baselines and AGCN-Mamba Model Optimizer Learning Rate Batch Size Weight Decay Learning Rate Schedule CNN-LSTM Adam 1e − 3 64 1e − 4 Step Decay (0.1/30 epochs) ST-GCN SGD (Momentum = 0.9) 1e − 1 64 5e − 4 Cosine Annealing Transformer AdamW 1e − 4 32 1e − 2 Warm-up + Cosine 2s-AGCN SGD (Momentum = 0.9) 1e − 1 64 5e − 4 Cosine Annealing ST-GCN++ AdamW 1e − 3 64 1e − 2 Cosine Annealing AGCN-Mamba AdamW 1e − 3 64 1e − 4 Cosine Annealing Open in a new tab Performance evaluation Comparison of Action Recognition Accuracy In order to objectively evaluate the overall performance of the AGCN-Mamba model in basketball action recognition and physical fitness evaluation, the study compared it with CNN-LSTM, ST-GCN, and Transformer models. Four key indicators are compared and considered: overall accuracy, training convergence, performance of each action category, and reasoning efficiency, as shown in Figs. 5 , 6 , 7 and 8 . Fig. 5. Open in a new tab Top-1 accuracy comparison of models on the test set Fig. 6. Open in a new tab Validation accuracy convergence curve during training Fig. 7. Open in a new tab F1-score radar chart for different basketball action categories Fig. 8. Open in a new tab Qualitative Comparison of Kinematic Quality Degradation Figure 5 shows the Top-1 recognition accuracy of different models on the test set. The proposed AGCN-Mamba model performs best with an accuracy rate of 94.8%. In contrast, the accuracy of the traditional CNN-LSTM model is only 88.6% due to the lack of explicit modeling of human skeleton topology. Although the accuracy of ST-GCN is 92.4% after introducing skeleton diagram structure, it is still not as good as the model designed in this study. Conclusion The advantages of the "adaptive adjacency matrix" in AGCN module in dealing with non-physical dependence such as the cooperation of the shooter's knee are verified again. The experimental results show that the accuracy of AGCN-Mamba is equal to or even better than that of high computational power Transformer. This fully shows that Mamba architecture is highly competitive while efficient computing. Based on the analysis of the model training process and the performance of each stage, the convergence curve in Fig. 7 shows that AGCN-Mamba architecture has obvious advantages in feature extraction ability. From the initial iteration, the accuracy of the model showed a trend of rapid increase, and it quickly entered the stable operation range at about the 40th round, and the overall performance fluctuated little, thus avoiding the common oscillation problem of the Transformer model. This proves that the structured prior of graph convolution and Mamba’s selective memory mechanism effectively accelerates the gradient optimization. Accordingly, the radar chart of a specific action in Fig. 7 further demonstrates the ability of the model to analyze complex motion patterns. Specifically, for actions that require a high degree of whole-body coordination, such as “shooting” and “defensive shovel”, the model achieves a peak F1 score of 0.952 and 0.941, respectively, which is significantly better than ST-GCN and CNN-LSTM. These results show that the model successfully captures the dynamic compensation characteristics caused by the dependence between non-physical limbs, and can accurately evaluate difficult technical actions, even if it shows similar accuracy in simpler action recognition tasks. To objectively evaluate the performance of AGCN-Mamba in basketball scenarios, this study compared it not only with the three main baselines introduced earlier but also included two additional state-of-the-art models. In total, five representative models were analyzed: CNN-LSTM, ST-GCN, Transformer, ST-GCN++ (Enhanced Spatial-Temporal Graph Convolutional Networks), and 2s-AGCN (Two-stream Adaptive Graph Convolutional Networks). All models were trained and tested under the subject-disjoint split protocol, ensuring no overlap of athletes between training and test sets (Table 6 ). Table 6. Performance metrics of baseline models and AGCN-Mamba on action recognition and kinematic quality regression Model Top-1 Accuracy (%) Kinematic Quality MAE Pearson Correlation ( r ) CNN-LSTM 88.59 0.095 0.81 ST-GCN 92.37 0.088 0.84 Transformer 94.48 0.068 0.87 2s-AGCN 93.15 0.078 0.85 ST-GCN++ 93.91 0.074 0.86 AGCN-Mamba 94.82 0.065 0.91 Open in a new tab As shown in Table 6 , AGCN-Mamba achieved a Top-1 accuracy of 94.82% on the BAR dataset, surpassing all baselines. Compared with high-performance models ST-GCN + + and 2s-AGCN, it improved recognition accuracy by approximately 0.9% and 1.6%, respectively. These results demonstrate that the integration of adaptive graph learning with the Mamba linear state-space module allows the network to capture subtle, non-physical inter-limb coordination patterns, providing robust performance even in complex competitive action sequences. For fatigue evaluation, the regression task was explicitly defined as Kinematic Quality Degradation Assessment. Beyond classification accuracy, the model’s regression performance was quantified using Mean Absolute Error (MAE) and Pearson correlation coefficient (r). AGCN-Mamba achieved the lowest MAE (0.065) and the highest correlation (0.91) among all models, indicating that the predicted kinematic degradation scores align closely with expert coaches’ assessments. Compared to the Transformer, AGCN-Mamba improved r by 4.5%, attributable to the selective scan mechanism, which enhances sensitivity to key motion points in long sequences. Computational efficiency and edge deployment potential To examine algorithmic efficiency and practical device applicability, this study measured inference latency, parameter count, and computational complexity (FLOPs) for each model on an RTX 3090 GPU. Note that the single-frame inference latency for AGCN-Mamba (12.08 ms) includes only the end-to-end model computation and excludes OpenPose keypoint extraction. As shown in Table 7 , while Transformer achieves high recognition accuracy, its large FLOPs (45.2G) and inference latency (45.12 ms) limit deployment on resource-constrained devices such as mobile GPUs. In contrast, AGCN-Mamba leverages the linear complexity of the state-space model, achieving significantly reduced parameter count (2.8 M) and computational cost (1.2G FLOPs), with inference efficiency roughly 15× higher than ST-GCN++. These results highlight the model’s potential for real-time kinematic analysis on edge devices, though further validation on embedded platforms such as Jetson or mobile GPUs is needed for practical deployment and energy-efficiency evaluation. Table 7. Comparison of computational efficiency and hardware costs Model Params (M) FLOPs (G) Inference Latency (ms) FPS CNN-LSTM 5.2 12.4 8.92 112.1 ST-GCN 3.1 18.5 18.45 54.2 Transformer 12.8 45.2 45.12 22.1 2s-AGCN 3.5 22.1 20.12 49.7 ST-GCN++ 3.2 19.2 19.30 51.8 AGCN-Mamba 2.8 1.2 12.08 82.7 Open in a new tab Ablation study In order to further analyze the internal mechanism of model performance improvement, Table 8 shows the ablation experimental results for spatial and temporal modules. The results show that each component brings significant performance improvement, which verifies the rationality of the proposed architecture. Specifically, when the standard GCN with fixed topology is replaced by AGCN (Variant A) with an adaptive adjacency matrix, the accuracy is improved by 1.89%, which shows that dynamic learning of non-physical connections is very important to capture the compensation features between limbs. When the traditional LSTM (variant B) is replaced by the Mamba module, the performance of the model is improved even more, reaching 2.53%, which highlights the advantages of the Mamba selective scanning mechanism in modeling the fitness characteristics of long sequences and its ability to extract key frame information more effectively. In the end, the integrated model, which combines AGCN and Mamba, achieves the best Top-1 accuracy of 94.82%. The above performance improvement (+ 3.67%) is not the result of simple linear superposition, but the result of the interaction between high-quality spatial features and time series models. The input of high-quality spatio-temporal data greatly enhances the reasoning ability of time series. Table 8. Ablation experiment results Model Variant Spatial Module Temporal Module Top-1 Accuracy (%) F1-Score Performance Gain Baseline GCN (Fixed Graph) LSTM 91.15 0.898 - Variant A AGCN (Adaptive) LSTM 93.04 0.919 1.89% Variant B GCN (Fixed Graph) Mamba 93.68 0.925 2.53% Full Model AGCN (Adaptive) Mamba 94.82 0.935 3.67% Open in a new tab To further verify the necessity of the multi-task learning (MTL) architecture, this study compared the performance of single-task models against the full MTL framework. As shown in Table 9 , a classification-only model (Action Only) achieved slightly lower Top-1 accuracy, while a regression-only model (Fitness Only) exhibited significantly higher MAE. These results demonstrate that the action recognition task provides essential semantic constraints for kinematic quality degradation assessment, while the quality regression task acts as auxiliary supervision, capturing subtle motion deformations and enhancing the classifier’s robustness. In other words, positive transfer occurs between tasks, improving overall predictive performance. Table 9. Ablation study on multi-task learning effectiveness Task Configuration Top-1 Accuracy (%) Kinematic Quality MAE Pearson Correlation ( r ) Action Only 93.15 – – Fitness Only – 0.088 0.82 Full MTL 94.82 0.065 0.91 Open in a new tab Cross-scenario generalization To assess the generalization and robustness of the AGCN-Mamba framework in skeletal modeling, this study conducted experiments on the NTU RGB + D 60 benchmark dataset using the Cross-Subject (X-Sub) protocol, where training and test sets comprise entirely different subjects. This rigorously evaluates the model’s ability to handle individual physiological differences. As shown in Table 10 , AGCN-Mamba achieved 91.25% accuracy on general action recognition, improving nearly 10% over the baseline ST-GCN and slightly outperforming the state-of-the-art ST-GCN++ (89.84%). These results demonstrate: Table 10. Cross-subject accuracy comparison on NTU RGB + D 60 Model CNN-LSTM ST-GCN Transformer 2s-AGCN ST-GCN++ AGCN-Mamba Accuracy (%) 78.35 81.50 90.12 88.52 89.84 91.25 Open in a new tab The precision of the AGCN spatial module in capturing interactions among a large number of nodes. The generalization of the Mamba temporal module for modeling long-range sequences, even in non-professional or daily-life actions (e.g., walking, social gestures). This cross-scenario performance confirms the framework’s capability to extract robust features that can be applied to the complex kinematic quality degradation assessment in the BAR dataset. Sensitivity analysis of label components The kinematic quality degradation label was constructed from three components: DTW deviation, variations in action duration, and center-of-mass (COM) stability. To investigate the contribution of each component to the model’s final predictive performance, a leave-one-out strategy was adopted to regenerate the labels and retrain the model. As reported in Table 11 , removing the DTW deviation resulted in the most pronounced decline in the correlation coefficient (r), which dropped from 0.91 to 0.82. This finding indicated that distortions in trajectory shape constituted a core feature for identifying fatigue-induced degradation in movement quality. In contrast, the contribution of COM stability was more prominent in actions characterized by large COM fluctuations, such as defensive footwork. Table 11. Sensitivity analysis results of label generation components Ablated Component Kinematic Quality MAE Pearson r Performance Drop None (Full Label) 0.065 0.91 – Remove DTW deviation 0.092 0.82 9.89% (largest) Remove action duration 0.078 0.86 5.49% Remove COM stability 0.072 0.88 3.30% Open in a new tab For qualitative validation, this study analyzed keypoint behavior during basketball shooting actions. As illustrated in Fig. 8 , comparing sequences of peak performance (score = 1.0) and degraded performance (score = 0.4): At the trajectory level (Fig. 8 a), wrist keypoints in peak states (blue solid lines) follow smooth paths, whereas degraded sequences (red dashed lines) exhibit noticeable deviations and instability toward the action end. At the velocity stability level (Fig. 8 b), red shadowed regions indicate high-frequency “micro-compensation” fluctuations caused by fatigue. These micro-compensation patterns reflect the athlete’s non-standard joint adjustments to maintain action goals under muscular fatigue. The results strongly support the effectiveness of the model’s regression branch, showing that AGCN-Mamba can extract subtle kinematic features from visual streams that traditional methods cannot detect. Discussion From the perspective of architectural design and long-sequence modeling, the proposed AGCN-Mamba framework effectively leveraged the strengths of linear SSMs to capture complex dynamical evolutions. Unlike SMGNFORMER proposed by Li et al. (2025) [ 28 ], which combined Mamba with graph transformers to enhance human pose estimation accuracy, SMGNFORMER still depended on relatively complex attention mechanisms. These mechanisms tended to introduce computational redundancy when modeling long-horizon motion streams. In contrast, the proposed framework employed AGCN to model non-physical coordination dependencies, such as the synergy between knee flexion and wrist motion during shooting actions. This design was further integrated with Mamba’s selective scanning mechanism, enabling deep feature extraction while maintaining a low parameter count. This behavior was consistent with the findings of Wang et al. (2025) [ 29 ] in text-controlled motion generation, where Mamba-based architectures exhibited superior linear scalability compared with conventional Transformers in temporal grounding and complex spatiotemporal alignment. These properties formed the core technical basis for capturing the kinematic degradation patterns defined in this study. With respect to computational efficiency, it was necessary to distinguish clearly between algorithmic efficiency and device-level efficiency. The inference latency of 12.08 ms and the low FLOPs reported in Table 6 primarily reflected the theoretical advantages of the proposed algorithm. As highlighted by Ahmad et al. (2025) in their study on byte-level latency optimization for Mamba models [ 30 ], practical edge deployment remained constrained by hardware instruction sets and memory bandwidth. Although the proposed model required substantially fewer computational resources than baseline methods, the reported latency only accounted for the core model inference. It did not include system-level overheads, such as OpenPose-based keypoint extraction or Kalman filtering during preprocessing. To address this discrepancy, the knowledge distillation strategies proposed by Ahmad et al. offered a promising direction for future optimization. By transferring the spatiotemporal reasoning capabilities of large-scale Mamba models into more compact byte-level latent representations, the gap between algorithmic complexity and real-time constraints on physical edge devices could be further reduced. Importantly, this study explicitly reframed its primary contribution from physiological fatigue prediction to kinematic quality degradation assessment, thereby improving the biomechanical rigor of the evaluation framework. Qualitative visualizations, such as those shown in Fig. 8 , indicated that sequences with lower degradation scores were consistently accompanied by pronounced micro-compensation fluctuations. Although these visually derived proxy indicators could not fully replace gold-standard physiological measurements, including heart rate or blood lactate concentration, they provided a non-invasive and real-time auxiliary decision-support tool suitable for high-intensity competitive environments. Future work will build upon multimodal alignment strategies, such as those proposed by Wang et al., to explore domain adaptation between visually derived kinematic degradation labels and multi-source physiological sensor data. This integration aims to establish a closed-loop mapping from observable visual degradation to underlying physiological fatigue, thereby further enhancing the applicability of the proposed framework in real-world edge deployment scenarios. Despite the use of AGCN-based adaptive modeling to partially compensate for the absence of depth information in monocular vision, single-camera pipelines may still suffer from coordinate drift during long-duration matches with frequent and large viewpoint changes. In future research, explicit depth constraints or geometric view augmentation techniques inspired by recent advances, such as DepthSSC [ 31 ] and HarmonicNeRF [ 32 ], will be incorporated. These approaches are expected to improve the spatiotemporal consistency of skeletal sequences in 3D space, enabling more robust and reliable monitoring of athletic performance. Conclusion This study integrated virtual sensing techniques with an AGCN-Mamba deep learning architecture to establish a non-intrusive framework for assessing kinematic quality degradation in basketball athletes. The proposed approach effectively addressed the inherent intrusiveness of conventional hardware-based sensors in competitive environments, as well as the computational bottlenecks associated with long-sequence motion analysis. Experimental results demonstrated that, under a subject-independent evaluation protocol, the model achieved an action recognition accuracy of 94.82%. Moreover, the predicted kinematic quality scores exhibited a strong correlation with expert assessments, reaching a coefficient of 0.91. Combined with an inference latency of only 12.08 ms and a lightweight computational cost of 1.2 GFLOPs, these results confirmed the model’s ability to balance high-precision modeling with real-time edge-side deployment requirements. Although the current framework remained constrained by the absence of depth information in monocular vision and susceptibility to extreme occlusion scenarios, the findings clearly validated the effectiveness of adaptive graph learning and linear state space scanning mechanisms. In particular, these components demonstrated a pronounced advantage in capturing latent inter-limb compensation patterns and long-term fatigue-related degradation trends. The kinematic indicators derived in this study provided a statistically grounded proxy feedback mechanism for sports training applications. More broadly, this work established a solid theoretical and technical foundation for the future development of low-power intelligent sports monitoring systems that integrate multi-source physiological signals and incorporate full 3D spatial perception capabilities. Acknowledgements Not applicable. Authors’ contributions Z.Y., Y.Z. wrote the main manuscript text, prepared figures, tables and equations. All authors reviewed the manuscript. Funding Not applicable. Data availability The data used to support the findings of this study are all in the manuscript. Declarations Ethics approval and consent to participate Not applicable. Consent for publication Not applicable. Competing interests The authors declare no competing interests. Footnotes Publisher’s note Springer Nature remains neutral with regard to jurisdictional claims in published maps and institutional affiliations. References 1. Vretaros A. Comparison of physical fitness level among different competition categories in women’s basketball: A systematic review. Turkish J Kinesiol. 2024;10(3):207–31. [ Google Scholar ] 2. Abed Saheb HF, Malih FA, Demirci N. The impact of a High-Fitness program on developing performance endurance for players of the basketball skills challenge competition. Mod Sport. 2024;23(1):157–66. [ Google Scholar ] 3. Hou Y, Li Z, Li H. Sensor based interactive digital entertainment and gamified training to alleviate basketball player fatigue. Entertainment Comput. 2025;52:100838. [ Google Scholar ] 4. Zhang J, Guo R, Zhu Y, Che Y, Zeng Y, Yu L, et al. Sensor-Driven Real-Time Recognition of Basketball Goal States Using IMU and Deep Learning. Sensors. 2025;25(12):3709. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 5. Han M, Fan L. A short-term energy consumption forecasting method for attention mechanisms based on spatio-temporal deep learning. Comput Electr Eng. 2024;114:109063. [ Google Scholar ] 6. Petrigna L, Pajaujiene S, Musumeci G. Physical fitness assessment in wheelchair basketball: A mini-review. Front Sports Act living. 2022;4:1035570. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 7. Ibáñez SJ, Piñar MI, García D, Mancha-Triguero D. Physical fitness as a predictor of performance during competition in professional women’s basketball players. Int J Environ Res Public Health. 2023;20(2):988. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 8. Cao S, Liu J, Wang Z, Geok SK. The effects of functional training on physical fitness and skill-related performance among basketball players: a systematic review. Front Physiol. 2024;15:1391394. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 9. Hita IPAD. Analysis of the role of basketball in improving physical fitness levels. Int J Educatio Elementaria Psychologia. 2024;1(2):89–97. [ Google Scholar ] 10. Xu Q, Silva RM, Zmijewski P, Li T, Li J, Yang L, Clemente F. Enhancing physical fitness using recreational soccer and basketball: A parallel-controlled 8-week study involving overweight and obese individuals, with consideration of sex-related interactions. Biology Sport. 2025;42(1):47–58. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 11. Liu J, Liu X, Qu M, Lyu T. Eitnet: An iot-enhanced framework for real-time basketball action recognition. Alexandria Eng J. 2025;110:567–78. [ Google Scholar ] 12. Khobdeh SB, Yamaghani MR, Sareshkeh SK. Basketball action recognition based on the combination of YOLO and a deep fuzzy LSTM network: SB Khobdeh et al. J Supercomputing. 2024;80(3):3528–53. [ Google Scholar ] 13. Zhang L. Deep learning based fine-grained recognition technology for basketball movements. Syst Soft Comput. 2024;6:200134. [ Google Scholar ] 14. Wang J, Zuo L, Cordente Martínez C. Basketball technique action recognition using 3D convolutional neural networks. Sci Rep. 2024;14(1):13156. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 15. Tanveer MH, Fatima Z, Zardari S, Guerra-Zubiaga D. An in-depth analysis of domain adaptation in computer and robotic vision. Appl Sci. 2023;13(23):12823. [ Google Scholar ] 16. Lu Q. Sports-ACtrans Net: research on multimodal robotic sports action recognition driven via ST-GCN. Front Neurorobotics. 2024;18:1443432. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 17. Zahra I, Wu Y, Alshehri M, Alqahtani Y, Aljuaid H, Jalal A, Liu H. Dynamic graph neural networks for UAV-based group activity recognition in structured team sports. Front Neurorobotics. 2025;19:1631998. [ DOI ] [ PMC free article ] [ PubMed ] [ Google Scholar ] 18. Wang X, Tang Z, Shao J, Robertson S, Gómez MÁ, Zhang S. Hooptransformer: advancing NBA offensive play recognition with self-supervised learning from player trajectories. Sports Med. 2024;54(10):2663–73. [ DOI ] [ PubMed ] [ Google Scholar ] 19. Wen J, Liu D, Zheng B. ActionMamba: Action Spatial–Temporal Aggregation Network Based on Mamba and GCN for Skeleton-Based Action Recognition. Electronics. 2025;14(18):3610. [ Google Scholar ] 20. Fazzari E, Romano D, Falchi F, Stefanini C. Selective state models are what you need for animal action recognition. Ecol Inf. 2025;85:102955. [ Google Scholar ] 21. Güney E, Bayilmiş C, Cakan B. An implementation of real-time traffic signs and road objects detection based on mobile GPU platforms. IEEE Access. 2022;10:86191–203. [ Google Scholar ] 22. Güney E, Bayilmis C. (2024, May). YOLOv5-Based driver behavior monitoring system for safer roads on jetson xavier NX. In International Conference on Advanced Engineering, Technology and Applications (pp. 339–350). Cham: Springer Nature Switzerland. 10.1007/978-3-031-70924-1_25. 23. Güney E, Agirtas G, Bayilmis C. MongoDB based real-time monitoring heart rate using websocket for remote healthcare. Jurnal Ilmiah Teknik Elektro Komputer dan Informatika. 2022;8(4):500. [ Google Scholar ] 24. Qiu ZX, Zhang HB, Deng WM, Du JX, Lei Q, Zhang GL. Effective skeleton topology and semantics-guided adaptive graph convolution network for action recognition. Visual Comput. 2023;39(5):2191–203. [ Google Scholar ] 25. Fu Z, Chen J, Jiang K, Wang S, Wen J, Yang M, Yang D. Traffic police 3D gesture recognition based on spatial–temporal fully adaptive graph convolutional network. IEEE Trans Intell Transp Syst. 2023;24(9):9518–31. [ Google Scholar ] 26. Zeng H, Li Y, Niu R, Yang C, Wen S. Enhancing spatiotemporal prediction through the integration of Mamba state space models and Diffusion Transformers. Knowl Based Syst. 2025;316:113347. [ Google Scholar ] 27. Liu X, Zhang C, Huang F, Xia S, Wang G, Zhang L. Vision mamba: A comprehensive survey and taxonomy. IEEE Trans Neural Networks Learn Syst. 2025;1–22. 10.1109/TNNLS.2025.3610435. [ DOI ] [ PubMed ] 28. Li Y, Wang Z, Niu W. SMGNFORMER: Fusion Mamba-graph transformer network for human pose estimation. IET Comput Vision. 2025;19(1):e12339. 29. Wang X, Kang Z, Mu Y. Text-controlled motion mamba: Text-instructed temporal grounding of human motion. IEEE Trans Image Process. 2025;34:7079–92. [ DOI ] [ PubMed ] [ Google Scholar ] 30. Ahmad M, Mazzara M, Distefano S, Khan AM. Byte latent mamba with state space and knowledge distillation for hyperspectral image classification. IEEE Trans Geosci Remote Sens. 2025;63:1–15. [ Google Scholar ] 31. Yao J, Zhang J, Pan X, Wu T, Xiao C. (2025, February). DepthSSC: Monocular 3D semantic scene completion via depth-spatial alignment and voxel adaptation. In 2025 IEEE/CVF Winter Conference on Applications of Computer Vision (WACV) (pp. 2154–2163). IEEE. 10.1109/WACV61041.2025.00216. 32. Pan X, Yao J, Kou H, Wu T, Xiao C. (2024, October). Harmonicnerf: Geometry-informed synthetic view augmentation for 3d scene reconstruction in driving scenarios. In Proceedings of the 32nd ACM International Conference on Multimedia (pp. 5987–5996). 10.1145/3664647.3680851. Associated Data This section collects any data citations, data availability statements, or supplementary materials included in this article. Data Availability Statement The data used to support the findings of this study are all in the manuscript. Articles from BMC Sports Science, Medicine and Rehabilitation are provided here courtesy of BMC ACTIONS View on publisher site PDF (3.5 MB) Cite Collections Permalink PERMALINK Copy RESOURCES Similar articles Cited by other articles Links to NCBI Databases Cite Copy Download .nbib .nbib Format: AMA APA MLA NLM Add to Collections Create a new collection Add to an existing collection Name your collection * Choose a collection Unable to load your collection due to an error Please try again Add Cancel Follow NCBI NCBI on X (formerly known as Twitter) NCBI on Facebook NCBI on LinkedIn NCBI on GitHub NCBI RSS feed Connect with NLM NLM on X (formerly known as Twitter) NLM on Facebook NLM on YouTube National Library of Medicine 8600 Rockville Pike Bethesda, MD 20894 Web Policies FOIA HHS Vulnerability Disclosure Help Accessibility Careers NLM NIH HHS USA.gov Back to Top