Deep Reinforcement Learning based Dynamic Object Detection and Tracking from a Moving Platform

Abstract

This article proposes Deep Reinforcement learning inspired image based visual servoing (DRL-IBVS) controller for aerial robots. The proposed DRL-IBVS controller uses monocular images to map target’s bounding box based errors to linear-velocity command to the robot for following the target. Deep learning based object detection algorithm YOLOv2 architecture is used to identify the target in the image. The noise in detection is filtered using multi-object Bayes filter, where the prediction model utilizes a correlation tracker for target prior estimation. Filtered bounding box output is fed to the Deep Deterministic Policy Gradient (DDPG) for performing the visual servoing. DDPG agent is trained separately using the pinhole camera model and the kinematic relation between two drones (follower and target). Gazebo-based simulation environment is used to successfully validate the controller.