Flying in Dynamic Scenes With Multitarget Velocimetry and Perception-Enhanced Planning
Han Chen, Chih‐Yung Wen, Fei Gao, Peng Lu
Abstract
Micro drones have been widely applied in applications, such as aerial cinematography and environment exploration. Due to the presence of numerous demanding scenes, such as crowded obstacles and dynamic objects, remotely controlling a drone poses a significant challenge for humans. As a result, there is an urgent need for highly autonomous flight capabilities. This article presents a fully autonomous flight system in a complex dynamic environment, showing satisfactory performance in real-world tests, and outperforms the state-of-the-art works in both dynamic object perception, and flight safety and efficiency. A lightweight but effective multiobject velocimetry based on a cross-correlation algorithm and local points feature is proposed, with a robust image-based object classifier as the front end. Also, we plan the flight trajectory considering the camera's field of view and the uncertainty in the dynamic object's constant velocity model. At last, we further explore the benefits of vehicle's active yaw control for improving perception quality and flight safety.