This is published under the terms of CC BY-NC.

Downloaded: 117 times

Abstract

In remote dynamic hand-gesture recognition, uncertainties in timing and distance of gesture occurrences, coupled with the subtle bodily perturbations induced by arm movements, pose substantial challenges to the accurate extraction of gesture features. In this paper, we propose a lightweight real-time gesture recognition system based on support vector machines. By analyzing the Doppler features of different motion states, a Doppler weighting factor was constructed to suppress bodily micro-motion interference in the range-time spectrum, and achieve foreground extraction of gesture signals concurrently. Furthermore, prior to the extraction of HOG features, we employ Gaussian filtering to suppress abrupt transitions and noise inherent in the gesture signals. This preprocessing significantly enhances the stability of feature extraction. Subsequently, the extracted features are input into an SVM for training and classification. Experimental results demonstrate that, for five distinct gestures exhibited in two different states –– standing and seated –– within a range of 1 to 5 meters, the recognition accuracy reaches 96%. This proves the feasibility of the proposed methodology, and its potential to realize real-time gesture recognition.