Real-Time Hand Gesture Recognition for IoT Devices Using FMCW mmWave Radar and Continuous Wavelet Transform

In recent years, Frequency-Modulated Continuous-Wave radar has gained significant interest in human-related applications, including hand gesture and activity recognition. Through a very detailed analysis of the existing literature, it was observed that most current methods rely on classical signal processing in the Fourier domain (discrete Fourier transform) combined with deep neural networks, and the studies primarily focus on optimizing the classifier to improve overall performance. In this paper, a different approach is proposed, concentrating on a novel stage of radar signal pre-processing. Instead of the traditional discrete Fourier transform, the Continuous Wavelet Transform is used, allowing for direct extraction of relevant features from the raw radar signal and the construction of a deep neural network for hand gesture recognition using Frequency-Modulated Continuous-Wave radars. Extensive experiments demonstrated that our approach systematically increases classification accuracy, particularly in the case of simpler models of deep neural networks, which, due to computational efficiency, are suitable for edge-device applications. The dataset collected included 20 participants performing five standardized gestures, each repeated twice, yielding 200 initial samples. By applying geometric transformations, the number of samples increased to 1000. The network achieved an accuracy of up to 99.87% using the Morlet wavelet and demonstrated good generalization capability on five new participants in laboratory conditions, with an average recognition accuracy ranging from 82% to 84%. These results confirm that combining Continuous Wavelet Transform processing with the proposed architecture of deep neural networks provides a lightweight, efficient, and accurate solution for real-time hand gesture recognition in practical human\u2013computer interaction applications.