Optimized in-vehicle multi person human body pose detection

Abstract

The number of Shared Autonomous Vehicles (SAV) will increase in the coming years. The absence of human driver will create a new paradigm for in-car safety. This paper addresses this problem by presenting an approach to estimate the human body pose inside a vehicle. We propose to use a customized version of the OpenPose framework, to perform the task of human body pose detection for the front passengers inside a vehicle. The OpenPose method was been evaluated with three different backbones: VGG19, MobileNetV1 and MobileNetV2, using different hyperparameters and ablation scenarios. Moreover, synthetic images were used, which simulate a depth sensor perspective from the center of the front seats. The dataset is comprised by images with 1 and 2 passengers, from 18 different subjects inside of 7 different vehicles, thus making a total of 45360 different images. The OpenPose method with the MobileNetV2 backbone showed the most efficient results between precision and performance, achieving a mean Average Precision (mAP) of 90%, Area Under ROC Curve (AUC) of 73%, and 0.0189 seconds per image (s/img).