Privacy-Preserving Pose Estimation for Human-Robot Interaction

Abstract

Pose estimation is a critical technique for non-verbal human-robot interactions. However, the presence of a camera in private spaces raises privacy concerns for humans and could lead to distrust of the robot. In this thesis, we propose a privacy-preserving camera system to realize pose estimation. We cover the camera with a translucent filter that protects privacy and then introduce an image enhancement module, which is an end-to-end neural network designed to improve pose estimation performance on the filtered images (shadow) images. We train the neural network on two datasets: a new filtered image dataset and a synthetic hazy image dataset. Both training methods improve the detection rate and precision of pose estimation. Based on our experiment, we conclude that our method can protect humans' privacy while detecting humans' pose information effectively.