Gesture Recognition System Based on Time-Frequency Point Density of sEMG

Abstract

Gesture recognition technology based on surface electromyography signal (sEMG) has important application value in human-computer interaction, medical rehabilitation, and other fields. It is usually realized by extracting the characteristics of different finger movements and then using machine learning or deep learning algorithms to classify and recognize them. This process involves complicated calculations, and few studies have achieved this purpose by combining different fingers’ flexion or relaxation. In this study, we designed a wearable acquisition system to collect sEMG with five fingers flexed/relaxed, and then combined with the movement status of the five fingers to identify different gestures. And in the process of detecting the movement status of the finger, focusing on the deficiency of traditional methods in using empirical thresholds to detect the sEMG active segment, we proposed an adaptive recognition method based on time-frequency point density. This method innovatively uses the time-frequency point density (TFPD) as the characteristic parameter of the sEMG, and then adaptively normalizes the feature extraction results in the interval [−1,1]. Finally, it uses a binary judgment method based on sliding windows to identify whether the active segment of the sEMG starts or ends, thus judging the flexion/relaxation state of the fingers. A large number of experimental results show that this method can realize quasi-real-time recognition within 0.5s, and the accuracy rate is nearly 100%. The influence of individual differences can be weakened through normalized positive and non-positive values. Therefore, it has strong self-adaptability. In addition, this method is very practical in the gesture recognition system.

Keywords

• Active segment
• detection
• gesture recognition
• self-adaptability
• surface electromyography
• time-frequency point density