In-Cabin Detection, Localization and Classification based on mmWave Radar with TinyML

This paper presents a real-time in-cabin detection, localization, and classification solution with a millimeter-wave (mmWave) radar system-on-chip (SoC), Calterah CAL60S344-AE, enabled by tiny machine learning (TinyML). The beam-range-Doppler (BRD) features are proposed to alleviate the impact of interference and facilitate adults and infants classification. We develop two lightweight convolutional neural networks (CNNs) with label-smoothing techniques to overcome noisy labels and augment model generality. Extensive experiments show that the implementation can achieve over 96% accuracy for detection and over 90% accuracy for localization and classification in the worst case, with only 76 KB model size and around 44 ms inference time. Comparative analyses also underscore the importance of model generality and robustness for real-world applications, and demonstrate that satisfactory performance can be attained even by using a simple CNN model.