CNN-Based In-Vehicle Occupant Sensing Using Millimeter-Wave Radar

In-vehicle occupant sensing seeks to perform three primary tasks: detect one or more occupants in a vehicle, localize a detected occupant to a seat, and classify the detected occupant. While in-vehicle occupancy detection is very reliable using millimeter-wave radar, occupant localization and classification per seat are more challenging due to limited angular resolution, ambient reflections, and multipath reflections within the vehicle. In this paper, we propose a novel deep-learning method for high-accuracy occupant seat localization and classification. Our deep learning model extracts spatial features through a 3D convolutional neural network (CNN) and temporal features through a recurrent neural network (RNN), showing considerable improvement in accuracy as more data becomes available over time. Testing our proposed deep learning model on new groups of participants unseen during training and validation, we demonstrate an accuracy of 95.5% for occupant seat localization and an accuracy of 85.1% when classifying as baby, child, or adult.