FFCSLT: A Deep Learning Model for Traffic Police Hand Gesture Recognition Using Surface Electromyographic Signals

Using surface electromyography (sEMG) signals for gesture recognition can significantly improve the effects of recognition. Therefore, this article proposed a CNN-SE-LSTM-TCN feature fusion network (FFCSLT) for traffic police gesture recognition based on the characteristics of sEMG signals. First, an acquisition system with six-channel sEMG sensors was developed for acquiring sEMG signals during human movement, and the dataset of hand gestures of traffic police (TPG) was constructed, which contains a total of 36000 sets of data. Then, a squeeze-and-excitation (SE) block with adaptive channel weighting was added on top of the depthwise separable convolutional network (DSCN) to enhance the spatial features between each channel in the FFCSLT network. Meanwhile, a temporal convolutional network (TCN) was integrated into the long short-term memory (LSTM) to extract additional temporal features in the FFCSLT network. Finally, the comparing experiments with other methods were taken on two datasets: the self-collected TPG dataset, widely used sEMG Sensor Data Ninapro DB1. The experimental results show that our model has an accuracy of 98.89% on the TPG dataset and 96.52% on the Ninapro DB1 dataset, which is 2.22% and 0.75% higher than suboptimal methods, respectively. To further validate the proposed network, we also performed a variety of ablation studies.