RGB-D Based Visual Navigation Using Direction Estimation Module.

Target-driven visual navigation without mapping works to solve navigation problems that given a target object, mobile robots can navigate to the target object. Recently, visual navigation has been researched and improved largely by learning-based methods. However, their methods lack depth information and spatial perception, using only single RGB images. To overcome these problems, two methods are presented in this paper. Firstly, we encode visual features of objects by dynamic graph convolutional network and extract 3D spatial features for objects by 3D geometry, a high level visual feature for agent to easily understand object relationship. Secondly, as human beings, they solve this problem in two steps, first exploring a new environment to find the target object and second planning a path to arrive. Inspired by the way of humans navigation, we propose direction estimation module (DEM) based on RGB-D images. DEM provides direction estimation of the target object to our learning model by a wheel odometry. Given a target object, first stage, our agent explores an unseen scene to detect the target object. Second stage, when detected the target object, we can estimate current location of the target object by 3D geometry, after that, each step of the agent, DEM will estimate new location of target object, and give direction information of the target object from a first-view image. It can guide our agent to navigate to the target object. Our experiment results outperforms the result of state of the art method in the artificial environment AI2-Thor.