REACTER: Perception-Informed Adaptive Tracker in Cluttered Environments

Abstract

This letter presents REACTER, a perception-informed tracking framework that enables quadrotors to track moving targets safely and adaptively without occlusion in cluttered environments. The core idea of the proposed approach is to dynamically adjust the tracking distance and observation angle based on the environmental complexity (EC). To this end, an obstacle distribution information structure (ODIs) is first constructed and incrementally updated, facilitating the prediction of the target's motion intention and the EC quantification. Then, an observation region matching mechanism that integrates visibility criteria and potential corners is introduced to guide viewpoint sampling. Finally, a novel perception-informed path searching scheme is developed, followed by efficient trajectory optimization within an elaborately designed tracking flight corridor. We benchmark our planner against three state-of-the-art tracking methods and demonstrate its superiority in both simulations and real-world experiments. To benefit the community, we release our code as an open-source package.