Deep Hand Gesture Recognition: A Wavelet Scattering Alternative to Convolutional Networks

Abstract

Hand gesture recognition (HGR) is crucial for improving human-computer interaction, aiding people with disabilities, and enhancing industrial efficiency. Radars are a popular choice for HGR, as they can detect hand gestures in various lighting conditions, through obstructions, with low latency, and without line-of-sight. Deep Convolutional Neural Networks (DCNN) are commonly used to analyze radar signals and recognize complex hand gestures. However, training DCNN models requires significant computational resources, making real-time applications challenging. This study proposes using Wavelet Scattering Transform (WST) as a feature extractor to replace DCNN, while relying on lightweight traditional classifiers for identifying the class of hand movement. WST is a non-linear signal representation that preserves high levels of discriminability while maintaining stability under time-warping deformations. To compare DCNN against WST, the study used a publicly available database of ultra-wideband (UWB) impulse radar gestures, collected from eight participants performing twelve hand gestures. The results showed that WST can achieve an average accuracy of 95% across all subjects, making it a reliable, computationally efficient, and accurate alternative to DCNN. This is the first research demonstrating the effectiveness of WST against DCNN for radar HGR applications (to the best of the authors’ knowledge).