Viewpoint Selection for DermDrone using Deep Reinforcement Learning

Abstract

This paper presents an RL-based method to improve the performance of real-time 3D human pose estimation as a positioning feedback for DermDrone which is a micro sized quadrotor designed MetaOptima to capture high resolution full body images for dermatology application. The camera viewpoint is identified as the key parameter in the accuracy of monocular 3D human pose estimation. We present a deep reinforcement learning based method for determining the best viewpoint given the flight trajectory. Our goal is to present a reliable and accurate positioning feedback for DermDrone using a 3D human pose estimation algorithm. DQN and its variants (Double DQN, and Dueling DQN) were employed and their performances were investigated by conducting several simulations. The results confirm that RL-based viewpoint selection improve the performance of 3D human pose estimation.