Lightweight Radar Signal Classification Using Wavelet Features and Machine Learning for Embedded Industrial Applications

This paper presents a novel and computationally efficient method for classifying moving road traffic objects using FMCW radar. The proposed approach operates directly on raw radar IQ signals, applying Continuous Wavelet Transform (CWT) and extracting simple statistical features (mean, standard deviation, and maximum) to form compact feature vectors. Ablation studies confirmed that while individual features provide limited classification capability, their combination yields the best performance. Classification is performed using lightweight algorithms such as Random Forest, SVM, or kNN, achieving high accuracy (94\u201395% for six object classes) while maintaining minimal computational overhead. Unlike image-based CNN methods, the approach eliminates time-consuming spectrogram generation, enabling fast training and prediction suitable for real-time and embedded applications. Comparisons with state-of-the-art deep learning approaches demonstrate that the proposed method achieves comparable accuracy with significantly lower computational cost. FMCW radar provides robustness against adverse weather and lighting conditions and enables detection independent of optical visibility. This work contributes to the development of efficient radar-based signal processing and pattern recognition techniques, offering a practical alternative to vision-based classification for intelligent traffic monitoring and autonomous systems.