Solving Pediatric Vehicular Heatstroke with Efficient Multi-Cascaded Convolutional Networks

Abstract

Pediatric Vehicular Heatstroke (PVH) is the situation where children suffer fatal injuries due to heatstroke after being forgotten in vehicles. It is a severe social problem: According to incomplete statistics, at least 864 children have died due to PVH since 1998 in the USA alone, and another 22 lost their lives in 2020. In this paper, we developed a machinelearning based embedded warning system that mitigates such tragedies. Specifically, we present our Children in Vehicles (CIV) dataset, where we collected 2,076 positive samples of children and 1,529 negative samples of empty car interiors. We then present the framework and training process of our multi-cascaded convolutional network architecture that can detect children with a 98% accuracy. Furthermore, we demonstrate the power of our novel curriculum learning method, which improved the classification accuracy of our facial age estimator from 46% to 62% and its F1 score from 0.66 to 0.91. We also deployed our complete pipeline onto an embedded platform to present its overall feasibility. Additionally, we open-sourced our code and dataset for others to use & experiment with.