Real-Time Radar Hand Motion Recognition Using a Hybrid CNN–SVM Framework With FPGA Acceleration

Radar-based hand gesture recognition (HGR) systems often suffer from challenges related to sensor interference, environmental clutter, and limited dataset size, which complicate the training of deep learning models. This paper presents an efficient radar-based HGR framework that integrates radar signal preprocessing, deep learning-based feature extraction, and hardware acceleration on a System-on-Chip (SoC) FPGA platform. To investigate the impact of radar data representation on classification performance, two experimental pipelines are explored: an image-based radar representation and a continuous-valued range–time map (CV-RTM). The preprocessing stage includes clutter suppression, squared-magnitude transformation, Gaussian filtering, and normalization to enhance gesture-related radar signatures. For the image-based pipeline, binarized radar gesture images are used to train a hybrid convolutional neural network– support vector machine (CNN–SVM) classifier. The proposed CNN–SVM model achieves a CNN training accuracy of 93.3% compared to 92.2% for a CNN–LSTM–SVM architecture, while reducing model parameters by 75.5% and decreasing SVM training and prediction times by 29.6% and 32.6%, respectively. For the CV-RTM pipeline, multiple deep learning configurations are evaluated under varying signal-to-noise ratio (SNR) conditions using a curriculum learning strategy to improve robustness in noisy radar environments. Experimental results demonstrate that the combined CNN–SVM model with curriculum learning achieves the most stable performance across SNR levels ranging from −20 dB to 10 dB. Finally, the optimized model is deployed on a Xilinx Kria KR260 FPGA platform using the Vitis-AI framework. The hardware implementation achieves real-time inference with significant improvements in execution speed and throughput, confirming the suitability of the proposed framework for embedded radar gesture recognition applications.