A Lightweight Hand-Gesture Recognition Network With Feature Fusion Prefiltering and FMCW Radar Spatial Angle Estimation
Jingxuan Chen,Shisheng Guo,Shuo Lv,Guolong Cui,Lingjiang Kong
DOI: https://doi.org/10.1109/jsen.2024.3432972
IF: 4.3
2024-09-03
IEEE Sensors Journal
Abstract:In dynamic hand-gesture recognition, the utilization of multidimensional features in conjunction with deep learning technique enables the precise recognition of intricate gestural movements. However, excessive input features inevitably introduce information redundancy, thereby engendering heightened network complexity. In this article, we propose a lightweight CNN + LSTM network with prefiltering mechanism to achieve accurate and robust hand-gesture recognition. The prefiltering network is designed to enhance the stability of feature extraction and reduce the complexity of entire network. Series of range-Doppler (RD) spectra and azimuth-pitch-angle (APA) spectra are employed to representing the spatiotemporal characteristics of hand-gesture movements, thereby increasing the information dimension of gesture features without complex inputs. Additionally, a novel loss function calculation approach is introduced, which effectively training the network to obtain an accurate classification at the end of a gesture. Finally, we recruited ten volunteers for acquisition of the hand-gesture dataset, employing the gesture data from seven of these individuals for network training, and subsequently evaluating its performance using gesture data from the remaining three users, achieving a classification accuracy as high as 99.65%.
engineering, electrical & electronic,instruments & instrumentation,physics, applied