Overcoming Core Challenges in Monocular Camera Drone Navigation of Unknown Maze

Typically, systems for autonomous monocular navigation of a drone in an indoor unknown maze of obstacles consists of three main components: 1) sensing-part which is usually Visual Simultaneous Localization and Mapping (VSLAM- computing a 3D-map of obstacles and pose-estimation of the drone), 2) the policy part (usually computing an obstacle-free path), and 3) the control part that executes the flight instructions. Though realizing these components (e.g. sensing=VSLAM and policy=freePath) yields a drone that works the resulting system will not be robust enough to cover all types of mazes. Thus, realizing these components is not enough to produce a robust working system. For example, when searching in a maze of rooms and doors, often there can be smooth surfaces (such as doors and white walls) that have no key points to detect. Hence, VSLAM (and other alternatives) will fail in building its 3D map and calculating the drone position. In this work, we describe a set of additional 'sub-problems' whose solution enables any monocular-based sensing+policy to work in general real-life scenarios. This is especially true for a lightweight drone limited to a simple monocular camera and using Raspberry Pi for control. We used a system whose sensing consists of an obstacle/free-space detector ODV [3] and the policy is based on the Zigzag algorithm [2]. The set of sub-problems we have found include: obtaining fast processing rates, Fast accurate pose estimation, Overcoming featureless areas such as white-walls, and obtaining smooth flight trajectories. The experiments we describe, fully show that the solution of these sub-problems allows the Zigzag algorithm to efficiently scan rooms even when featureless surfaces exist.