Light Monocular Camera-based Obstacle Detection and Avoidance Algorithm for Small drone Flying in an unknown Maze

We consider the problem of a drone (quadcopter) that needs to detect and avoid an obstacle when autonomously flying in an unknown maze of walls and obstacles. The detector performs two tasks: • Detect that the drone is close to colliding with an object if the drone continues flying in its current direction. • Once a close object has been detected the drone will start turning and the detector should signal when to stop this turn as an obstacle-free direction has been detected. This camera-based detector must be of low complexity as it is intended to be used by low-weight/battery drones whose controllers (such as the Raspberry PI0) can not support heavy computations. This ability is a fundamental problem in the area of drones (even in general robotics). In this work, we propose such a camera-based detector whose complexity improves upon previously proposed techniques. In previous works, key points are extracted, and based on their relative position-disparity in two consecutive frames, either true distances or the shape of a close object are computed. Rather than compute distances or the shape of a close object we use the horizontal disparity (roll angle 0) just to compute the probability that there is a close object in front of the drone. This reduces the overall complexity significantly, yet our experiments in flight simulation mode show a high detection ratio.