Generation of Recovery Data to Re-localize a Visual SLAM Robot using Deep Learning

Abstract

Track Loss is a huge problem that occurs during a SLAM process which disturbs the localization and mapping process and it makes the robot dependent on human supervision. In an ideal scenario, a robot –while performing Visual SLAM- should be able to move in an unknown environment, prevent obstacles and map each and every corner of the room. But in real cases, even with perfect obstacle avoidance, there is a chance that the robot may not map the environment perfectly due to factors such as Camera Blur, Sudden rotation or feature-less images. In these cases, track is lost since image SIFT features are not matched frame-by-frame which is a requirement of running SLAM without interruption. As soon as track is lost, the mapping process is disrupted, even though the robotic agent doesn’t know, it keeps on moving which is problematic. Now unless the agent does not come back to a point from where an image has been made part of the map before, mapping remains halted and the agent is “lost”. Our Research is to provide the agent with “recovery data” as soon as it becomes lost. As soon as track is lost, the last image of the recorded map and the existing image extracted from the stopped robot are fed to a pre-trained classifier. This classifier decides the movement the agent must do in order to continue mapping. In order to demonstrate our proof of concept, we developed a novel tool by preparing an integrated set up of MINOS Simulator and ORB SLAM, the agent is able to maneuver over a simulated environment autonomously while performing SLAM simultaneously. In case of any track loss, our classifier guides MINOS to recover its tracks.