Discrimination Between Car Trunk and Child Motion Using Time-Domain Kurtosis and Skewness Features for In-Vehicle Child Presence Detection

Abstract

While developing child-presence detection with ultra-wideband radar, we observed that trunk-lid motion can mimic in-cabin child activity and cause false alarms. We address this issue with a lightweight, purely time-domain detector that operates on sliding-window skewness and kurtosis of the radar amplitude stream. Trunk operations give rise to extended, low-variation plateaus in these statistics, whereas child motion remains non-stationary, so that a simple rule based on synchronized stable runs can separate the two classes without learning-based models. The detector is implemented in a streaming fashion using incremental updates, leading to constant per-sample complexity, fixed decision latency and modest memory usage suitable for embedded electronic control units. We evaluate the method on data collected in real in-vehicle environments with diverse trunk operations and child activities, and the results show that the proposed criterion reliably suppresses trunk-induced false alarms without introducing misclassifications. Because it uses only distributional time-domain features from a single multi-channel ultra-wideband radar, the approach is privacy-preserving and readily integrable into existing child-presence detection pipelines.