This paper proposes a gesture recognition method based on drum shaped friction nanogenerator (DS-TENG) array, temporal convolutional feature compression (TCF), and Bayesian optimized random forest (BO-RF) classification model. Through a 2 × 2 DS-TENG array, we can effectively collect the superficial tendon vibration signals containing gesture information in the wrist. By using the TCF strategy, the 1D temporal signals of corresponding gestures are mapped into 4-channel 32 × 32 grayscale pixel images and compressed into 1D feature vectors for spatiotemporal features extraction. Subsequently, we correlated gesture information with feature vectors fed and the dataset into the trained BO-RF classification model, achieving a gesture recognition accuracy of 97.86%. Compared with traditional classification methods, the using of self-powered DS-TENG arrays effectively improves the measurement sensitivity of superficial tendon vibration signals, while the application of TCF strategy and Bayesian optimization method reduce overfitting risk and computational complexity during the model training and improve the gesture recognition accuracy of models effectively. In short, this work provides an efficient solution for gesture recognition in low-power wearable devices and has broad application potential in virtual reality, augmented reality, and medical rehabilitation fields.