Dynamic Obstacle Avoidance for an MAV Using Optimization-Based Trajectory Prediction with a Monocular Camera

Abstract

Vision-based algorithms are widely applied to micro-Air vehicles (MAVs) because of their limited takeoff weight. Conventional stereo camera requires a large baseline for long-distance detection, which is difficult for MAVs. The rapidly developing, learning-based, monocular depth estimation method can handle these problems and succeed remarkably well in providing acceptable depth in indoor (maximum distance: 10 m) and outdoor (maximum distance: 80 m) environments. For the safety of an MAV in outdoor environments, we, therefore, propose a monocular-camera-based dynamic avoidance system, along with obstacle motion estimation by depth estimation methods using the Kalman filter. To handle the position uncertainty of a dynamic obstacle and predict its future movement, a polynomial-fitting-based trajectory prediction method with a defined uncertainty range has been used. Subsequently, using quadratic programming (QP), a safe, corridor-based, spatiotemporal trajectory generation method is proposed to ensure the safety of the MAV. We validate the performance of our algorithm through simulation and real-world experiments using an MAV.