Implementation of C4.5 decision tree in Human Gesture Recognition based on Doppler radars

In order to improve the accuracy of gesture recognition, we investigate the feasibility of using a three-dimensional Doppler-radar array at 24GHz to recognize human gestures with a model consisted of ten classical gestures. On the basis of C4.5 algorithm, phase difference and spectral energy, extracted by correlation processing and power integral, are used as the features to construct decision tree and separate ten gestures. The experiment result shows that this system could achieve a high accuracy of classification reached 99.25%.

I. Introduction

In recent years, hand gesture recognition on mobile devices attracts various aspects of research. Based on optical cameras and image-processing algorithms, hand gesture recognition for advanced human-machine interface has been extensively studied in the area of computer science [1]. However, it is elusive to use camera-based hand gesture recognition solutions on ordinary occasions due to the large amount of computation and the sensitivity to ambient light. Meanwhile, Doppler-radar sensors show promising performance in non-contact applications such as human vital signs detection, structural health monitoring, or through-the wall life detection [2]. In addition, Doppler radars can be implemented in low-cost devices with a simple front-end heterodyned architecture where the required sampling rate of the analog-to-digital converter (ADC) is low and low-power low-cost embedded digital signal processing is possible. In 2016, Google ATAP lab announced their milestone work "Soli" which investigated using a 60GHz FMCW radar to acquire Frequency-Doppler maps of different hand gesture and fulfill the gesture recognition [3], [4]. However, the special chip adopted by this system is not procurable into wide applications at the moment. In terms of data processing, [5] proposes constructing hand gesture classification system with decision tree method and artificial supervised thresholds, whereas the supervised thresholds keeps weak robustness and complexity.