semg finger gesture recognition

This study addresses the challenge of deploying hand gesture recognition systems on low-power embedded devices, where existing approaches relying on multi-channel electromyographic (EMG) signals and complex models are impractical. For the first time, it systematically demonstrates the feasibility of achieving high-accuracy classification of ten distinct gestures using only a single-channel surface EMG (sEMG) signal combined with lightweight machine learning. By integrating multi-domain features—including time-domain, frequency-domain, higher-order zero-crossing rate, and relative intensity—and applying Pearson correlation-based feature selection followed by PCA or LDA for dimensionality reduction, the authors evaluate classification performance using neural networks (NN), k-nearest neighbors (KNN), and support vector machines (SVM). The optimal configuration achieves 90% accuracy, substantially reducing both hardware requirements and computational overhead, thereby offering an effective solution for low-cost, low-power gesture-based interaction.

This study addresses the challenge of unsupervised clustering and classification of surface electromyographic (sEMG) signals during functional upper-limb movements. The authors propose a four-stage data-driven pipeline comprising signal preprocessing, multi-criteria feature selection, Mahalanobis distance–based hierarchical clustering for gesture selection, and classification evaluation. By integrating Hilbert envelope analysis, mutual information, principal component analysis (PCA), and decision tree–based feature importance, they identify six representative gestures from the NINAPRO DB4 dataset that exhibit both biomechanical diversity and physiological plausibility, and determine a 200 ms window as optimal for temporal segmentation. Experimental results demonstrate that Extra Trees and artificial neural network classifiers achieve superior performance, confirming the method’s effectiveness and scalability for low-latency prosthetic control.

This work proposes a novel approach to high-accuracy, low-latency hand gesture recognition for real-time control in intelligent prosthetics and augmented reality by modeling surface electromyographic (sEMG) signals from the forearm as graph structures that capture muscle activation patterns. For the first time, graph neural networks are leveraged for efficient gesture classification, integrating sEMG signal processing, dynamic graph construction, and lightweight graph-based inference. Evaluated on eight subjects, the method achieves an average recognition accuracy of 99% with an end-to-end processing latency of only 48 ms on an M1 Pro CPU, substantially meeting the stringent demands of real-time interaction. These results underscore the innovative value and practical potential of graph-structured modeling in sEMG-based gesture recognition.

This study addresses poor generalization in non-invasive surface electromyography (sEMG) gesture decoding, primarily caused by inter-subject and intra-subject signal variability. We propose modeling multi-channel forearm sEMG signals as covariance matrices embedded in the Symmetric Positive-Definite (SPD) Riemannian manifold—thereby explicitly characterizing the spatial synergy topology of sEMG for the first time. Within this geometric framework, unsupervised clustering and supervised classification are jointly realized without domain adaptation or subject-specific calibration. The method is computationally efficient and highly robust, achieving significant improvements over state-of-the-art baselines on mainstream benchmarks—particularly yielding a 12.3% average accuracy gain in cross-subject gesture recognition. Our core contribution is establishing an sEMG-to-SPD manifold mapping paradigm, offering a geometric learning approach to myoelectric decoding that is both interpretable and strongly generalizable.

Current sEMG-based prosthetic hand control systems rely on synchronized force/motion sensors, increasing system complexity and hindering clinical deployment. This paper proposes a sensor-free regression-based control framework: using only two calibration points—maximum flexion and extension effort (±1)—along with their corresponding sEMG signals, it constructs a lightweight, near-linear model to decode continuous finger force magnitude and direction, and enable real-time classification of three hand gestures. We introduce the novel “two-point calibration, zero-sensor calibration” paradigm, overcoming the bottleneck of multimodal synchronous sensing. Comparative evaluations across Dendritic Net, Linear Net, MLP, and CNN show Dendritic Net and Linear Net achieve superior offline performance in force-direction classification and intermediate-force interpolation. Online experiments confirm high accuracy, low-latency gesture transitions, and smooth continuous force modulation—significantly enhancing clinical practicality and user-friendliness.

This study investigates the impact of electrode configuration on gesture decoding performance when using surface electromyography (sEMG) from the wrist for thumb gesture recognition—a factor not yet well understood. The authors systematically compare high-density and low-density sEMG setups, examining electrode placement, referencing schemes, channel count, and spatial density. Results demonstrate that electrode placement over the extensor side significantly outperforms the flexor side (HD accuracy: 0.871 vs. 0.821), unipolar referencing yields better performance than bipolar, and a 15-channel unipolar configuration achieves an accuracy of 0.885. Further increases in channel count yield diminishing returns. The findings highlight the necessity of balancing coverage and compactness in electrode layout and underscore that thoughtful configuration design is more critical than simply increasing electrode number, offering practical optimization strategies for wearable wrist sEMG systems.

This study addresses the performance degradation of existing surface electromyography (sEMG)-based gesture recognition methods as the number of gestures increases, their limited capacity to model long-range temporal dependencies, and the lack of effective fusion with other physiological modalities such as inertial and eye-tracking signals. To overcome these challenges, this work proposes EMG-CrossFormer, an end-to-end hybrid convolutional–Transformer architecture that, for the first time, integrates learnable gesture queries into a Transformer framework for multimodal sEMG recognition. The model employs cascaded cross-attention mechanisms to jointly capture local and global spatiotemporal features and supports flexible input from arbitrary modalities. Experiments demonstrate that EMG-CrossFormer achieves 79.16% accuracy using sEMG alone across four NinaPro datasets, which further improves to 92.79% when inertial signals are incorporated—significantly outperforming six state-of-the-art baselines.

This study addresses the significant performance degradation in multi-sensor surface electromyography (sEMG) systems caused by the failure of individual sensors. To enhance system robustness, the authors propose a fault-tolerant framework based on the maximum Fisher Discriminant Ratio (FDR). By combining handcrafted feature extraction with FDR analysis, the approach ranks sensor importance and quantifies each channel’s contribution to a rock–paper–scissors gesture recognition task through systematic sensor ablation experiments using a multilayer perceptron classifier. The method effectively distinguishes between critical and replaceable sensors, thereby demonstrating the feasibility of the proposed fault-tolerant mechanism. These findings offer valuable theoretical support for designing redundancy strategies in high-robustness sEMG devices and inform their clinical deployment.

This study addresses the limited autonomy experienced by individuals with upper-limb motor impairments during daily tasks by proposing a surface electromyography (sEMG)-based real-time teleoperation control framework. The system employs four-channel sEMG signals segmented via sliding windows and integrates a one-dimensional convolutional neural network with a hybrid approach combining threshold-triggered detection and a two-stage classification strategy to enable end-to-end gesture recognition and robotic arm control. Designed within a unified architecture, the method jointly optimizes accuracy and latency, achieving stable and reliable discrete gesture control on both simulated and physical robotic platforms. Experimental results demonstrate an average classification accuracy exceeding 90% with a latency of approximately 0.32 seconds.

This study addresses the challenge of distinguishing between premeditated and spontaneous hand gestures in real-time multi-user interaction scenarios using electromyographic (EMG) signals. Employing dual-channel surface EMG from the forearm during a “rock–paper–scissors” task, the work analyzes temporal characteristics of gesture execution and demonstrates for the first time that user intent can be detected from EMG signals up to 800 milliseconds before visible movement onset. Furthermore, it reveals that reactive EMG responses in an opponent’s muscles contain discernible information about the observed gesture. Through temporal analysis of EMG onset and peak features combined with machine learning models, the system achieves 63.4% accuracy in identifying premeditated gestures and 53.6% when generalized to spontaneous gestures. Notably, gesture recognition from the opponent’s EMG peaks reaches 65% accuracy, albeit lagging visual action by 2082 milliseconds, thereby opening a novel pathway for interactive intention recognition.