Classification of Hand Motions Using Spatial Information in HDEMG Signals with HOG Features

Wearable assistive robotic systems require insight into users' motion intentions to provide intuitive assistance. Bio-signal modalities such as EEG, fNIRS, or sEMG can intercept signals from the nervous system, accessing information related to intended motions. However, technical challenges persist in interpreting the acquired information from these modalities, especially when dealing with a larger number of motions. In such cases, High-Density Electromyography (HDEMG) can offer measurements from a higher number of channels, providing more comprehensive information essential for motion classification. This study proposes a method to comprehend the information contained in HDEMG by analysing spatial changes in muscle activation during various motions. It utilizes histogram of gradient features derived from heatmaps associated with muscle activation. The proposed method aims to classify 12 different hand motions using a support vector machine-based classifier. Results demonstrate an average classification accuracy of 95% through 5-fold cross-validation involving 8 subjects. The high accuracy showcases the effectiveness of utilizing spatial variations in muscle activity to estimate human motion intention using HDEMG-based methods, particularly in potential robotic applications.