Abstract

This paper presents a method for detecting infants in passenger vehicles using a millimeter-wave FMCW-MIMO radar system with a concise yet interpretable Approximate Decision Tree (ADT) model. Unlike conventional approaches that rely on approximately 70 features and opaque machine learning classifiers, the proposed method reduces the feature set to about 20 by applying Constant False Alarm Rate (CFAR) and clustering algorithms to radar point-cloud data, and then consolidating the most effective splitting rules extracted from an XGBoost ensemble into a single ADT. This feature reduction not only improves interpretability but also preserves high accuracy, achieving a missed-detection rate of 0.6% compared with 1.8% in baseline experiments. The proposed pipeline allows transparent model visualization, enabling differentiation between infants and other objects even in challenging scenarios with moving items in empty seats. These results demonstrate that the method achieves both strong predictive performance and model interpretability, making it well-suited for safety-critical automotive applications.

Pages (from-to)	35958-35970
Volume	14
Publication status	Published - 2026