Monocular Gate Perception Based on ICP-like Optimization for Autonomous Drone Racing

Autonomous drone racing is a challenging task. It requires UAVs to autonomously cross an obstacle gate in space with unknown pose only by relying on onboard sensors and computing devices. Therefore, reliable detection and pose estimation of the gate is crucial. The accuracy of gate perception directly determines the success rate of UAVs in crossing the obstacle gate. In this paper, a monocular gate perception method based on ICP-like optimization is proposed. It mainly includes a frontend obstacle gate detector based on ellipse fitting and a backend pose optimization estimator based on ICP-like optimization. And a two-stage gate pose estimation method is introduced, which estimates the coarse gate position based on the detection results as the initial value to guide the nonlinear optimization of the pose estimator. We have conducted sufficient qualitative and quantitative experiments in both outdoor simulator scene and indoor real scene. In a large number of tests, our method demonstrates fast and accurate target detection and pose estimation performance and strong environmental adaptability. The real flight experiments prove that the method can well meet the needs of gate perception in autonomous drone competitions.