Movement and Gesture Recognition Using Deep Learning Technology

Abstract:

For several decades, the pattern recognition of movement and gesture shows promise for human-machine interaction in many areas. A remarkable application in this area is gesture recognition for upper limb amputees using surface electromyography (sEMG) to capture the muscle activation as electrical signals. Another well-known application of in this field is human activity recognition (HAR). Most HAR applications are based on raw sensor inputs such as accelerometer and gyroscope signals which show its ability in learning profound knowledge about movement recognition [1]. Within the field of signal-based gesture recognition, traditional machine learning (ML) approaches have been widely used [2]. ML models give a high accuracy with large amounts of hand-crafted, structured, and under controlled data. However, traditional ML models require lengthy offline and batch training which is not incremental or interactive for real time application. In addition, ML models always cost a long period of time to extract a set of reliable features especially for high-dimensional, complex and noisy data because of the various situations in practical applications. Besides the ML methodologies, in recent years, the use of deep learning (DL) algorithms has become increasingly more prominent for their tremendous ability to extract and learn features from large amounts of data [3]. Compared to ML models, DL models make it possible for artificial intelligence to train the networks without hand-craft feature extracting. The aim of this work is to develop DL based methods for human movement and gesture recognition from time-series signals such as obtained using sEMG and IMU signals. We would like to understand the performance of DL for time-series signal analysis and accuracy, as to our knowledge, this aspect is still understudied. A series of experiments have been conducted to achieve it with different datasets and signals. The DB1 is a HAR dataset from the UCI repository. The DB2 and DB3 are sub- datasets of Ninapro database contains the recordings of 17 gestures from subjects by collecting sEMG signal. There are 4 different DL models designed for the experiments to find out the optimum solution by performance comparison: a 1-D CNN, a LSTM model, a C-RNN and 3+3 C-RNN. This is an extended abstract of a poster for the conference. The details of datasets and models are described in the methodology section, followed with the result section to present the results of different DL models on datasets.

Download PDF