A Semantics-Driven Approach for Safe and Reliable Obstacle Avoidance of UAVs

Abstract

Autonomous obstacle avoidance for UAVs in complex low-altitude dynamic environments remains challenging due to the dual difficulties of perception ambiguity and the absence of semantic risk assessment. Existing geometric-based approaches fail to distinguish the potential risks of heterogeneous obstacles, while the instability of lightweight detectors on resource-constrained platforms further limits practical deployment. To address these challenges, this letter presents a semantics-driven safe and reliable obstacle avoidance system. We first design a probabilistic semantic-geometric fusion perception module. By utilizing semantic-geometric fusion tracking and recursive Bayesian filtering, this module eliminates association ambiguity and detection flickering, ensuring perception stability. Building on this, a semantics-driven MPC planner is proposed. By constructing a heterogeneous risk cost function and a dynamic soft constraint mechanism, the system achieves differentiated risk avoidance for various obstacles. Extensive simulations and real-world deployment on UAV platforms demonstrate that the proposed method significantly reduces the high-risk collision ratio while maintaining a high success rate, thereby enhancing flight safety in human-robot coexistence scenarios.