Novel Radar based In-Vehicle Occupant Detection Using Convolutional Neural Networks

Abstract

Vehicle Occupant Detection has gathered attention with the advancement of Connected Automated Vehicles (CAVs) since it enhances vehicular safety features and contributes to Vehicle-to-Everything (V2X) communication features. In this paper, a novel Frequency Modulated Continuous Wave (FMCW) radar-based occupancy detection utilizing Convolutional Neural Networks (CNN) is introduced. The proposed methodology tackles disadvantages posed by visual and sensor-based methods when privacy, computational complexity, line-of-sight requirements, and robustness are concerned. The system uses time-domain raw radar data signals to form visual heatmaps based on signal intensity variation caused by presence of a target. The heatmaps developed for each data frame acts as an input to the neural network. Visually generated signal based heatmaps differentiate three classes of vehicle occupancy: vacant, driver seat and rear passenger occupancy. The adapted CNN architecture is an implementation of transfer learning where a version of the VGG-16 pretrained model consisting of 16 convolutional layers is used. A validation accuracy of 96.88% is achieved with a dataset containing 1000 heatmap images for each class. The results conclude that radar generated time domain heatmaps efficiently detect vehicle occupancy employing transfer learning even with smaller datasets.