Multi-Class Object Recognition in Autonomous Driving Using Radar Heatmaps and Lightweight Machine Learning

mmWave radar plays a critical role in autonomous vehicles' perception due to its robustness under varying weather and lighting conditions. However, converting raw radar signals into usable object representations is challenging because of noise and computational constraints. This study presents a radar-based object classification framework using range-azimuth and range-Doppler heatmaps generated from raw ADC data of a 77GHz FMCW radar. The heatmaps are used to train two individual linear Support Vector Machine (SVM) classifiers for multi-class object detection. The models are evaluated using precision, recall, F1-score, confusion matrices, and ROC curves. Experimental testing shows that the range-azimuth model achieves a precision of 0.94, as compared to the range-Doppler model, which achieved 0.89. The results confirm that classical lightweight machine learning techniques offer a strong balance between accuracy and computational efficiency in automotive radar perception systems.