A guideline for gesture recognition developers with triboelectric gloves: Key insights on handcrafted feature extraction approach
Thiago Simões Dias, José Jair Alves Mendes, Heitor Silvério Lopes, Thiago H. Silva, Sérgio Francisco Pichorim · Biomedical Signal Processing and Control · 2025
Sign language recognition systems, particularly wearable system-based ones, are valuable tools that can enhance communication for hearing-impaired people. In this field of research, deep learning approaches have been widely explored for gesture classification; however, they require large datasets and substantial computational resources. This study investigates an approach for automatic segmentation and handcrafted feature extraction for gesture recognition using triboelectric glove signals. The aim is to provide key insights into the triboelectric glove system, which can serve as a guideline for developing gesture recognition devices. This work utilized a public database that includes 50 isolated words from American Sign Language. Three feature domains were analyzed (temporal, spectral, and statistical), and feature selection was carried out using the Fast Correlation Based Filter (FCBF) method. For classification, different models were tested, including k-Nearest Neighbors (KNN) and Random Forest (RF). The use of temporal features reduced the amount of information required for classification. When employing all feature domains, fewer sensors were required for classification. As left-hand gestures exhibit a lower probability of occurrence (i.e., higher entropy), they provide more critical information for accurate classification. Consequently, the model requires more sensors on the left hand than on the right hand. For future glove designs, considering the employment of all feature domains, sensors should be placed only on the index fingertip, as well as the index and middle fingers of the right hand. This approach can lead to lower costs and reduced prediction time for the system.