Evaluating Tracking Failure in Omni SLAM using Virtual Environment

Abstract

When training an agent for either path planning or Simultaneous Localization and Mapping (SLAM), datasets that include various scenarios according to the agent’s physical limitations are required. Although there are a variety of datasets available, training is limited to the perspectives provided by the datasets given. To combat this, we have used a virtual environment called Habitat Simulator [30, 25, 17]; a virtual space that can load various indoor places which resemble real-life places and where the agents can be trained for path planning or object retrieval. We have used these datasets to capture omnidirectional images; top, bottom, front, back, left, and right views for an agent, generated a cube map and finally converted it to fisheye images with a Field of View (FOV) of 360-degrees. These datasets are then used on ORB-SLAM3 [23], a widely used SLAM algorithm, to determine the SLAM’s performance on said datasets to see where the SLAM succeeds and where it encounters challenges.