High Dimensional Computing Gesture Recognition

Overview

The primary objective of this study is the Improvement of gesture recognition and classification accuracy through the use of the HDC algorithm compared to other classification methods (KNN, RF, SGD, NC). The recognition rate will be expressed by the sensitivity and specificity of gesture recognition. The model will be trained on a portion of the dataset and tested on the remaining part to avoid any bias. The secondaries objectives are the : * Improvement of gesture recognition accuracy with our HDC algorithm compared to other standard models. * Calculation of gesture recognition rates depending on the number of electrodes used and their position. * Subject's assessment of device comfort rated above 6 on a 10-level visual analog scale. * Subject's assessment of ease of performing the gesture rated above 6 on a 10-level visual analog scale.

This project aims to work on gesture recognition based on surface electromyography (EMG) recorded on the forearm. The CEA is currently developing a learning algorithm based on hyperdimensional computing designed to improve the accuracy and latency of gesture recognition. Unlike conventional computing methods, the developed approach relies on (pseudo) random hypervectors. This brings significant advantages: a simple algorithm with a well-defined set of arithmetic operations, extremely robust to noise and errors, with fast, one-pass learning that could ultimately benefit from a memory-centric architecture with a high degree of parallelism. This research could lead to multiple applications, such as video gaming or the metaverse, but also strongly interests the healthcare field, for example in robotic prostheses, tele-surgery applications, or simply medical training using virtual reality applications.