TROVE feature based visual-SLAM for environment rich in man-made structures

Abstract

For any robot and vehicle to perform autonomous navigation in an unknown environment, ego-pose perception and environment mapping is pivotal. Simultaneous localisation and mapping (SLAM) is an extensively studied subject serving that purpose. A wide range of sensors can be utilised in a SLAM system: an inertial measurement unit (IMU), GPS, ultrasonic sensor, LiDAR, RGB-camera and depth-camera. Among them, visual cameras are of major interest for their being economical and rich in information. This work investigates the problem particularly in an indoor environment by introducing a novel type of image feature which is associated with a 3D counterpart. The proposed method not only addresses the problem of localisation but also delivers the mapping in a more semantic form. Through this work, the candidate intends to provide a new perspective to visual-SLAM approaches.

The proposed visual-SLAM method centres around the novel image feature: TROVE (Three-rays-and-one-vertex). By observing an environment that is rich in man-made structures, one can often discover many regularities comprising of horizontal and vertical edges at the boundaries of objects and buildings. Those regularities can be distinctive and descriptive landmarks for navigation. The TROVE feature originates from the projection of those structures in an image. This work has developed a series of tools to detect the feature and a complete SLAM system based on it. Experiments have been conducted with a stereo camera to evaluate the feasibility and accuracy of the proposed method. In comparison with other state-of-the-art methods, the proposed one is lightweight but has high accuracy while producing a reusable and more intuitive map.