Side-Track: Pose-Aware Aerial Tracking with Semantic-Aware Anisotropic Planning

Autonomous aerial tracking of dynamic human targets is essential for applications like aerial cinematography and human-robot interaction. While state-of-the-art trackers have effectively addressed flight safety and target visibility in cluttered environments, they typically treat the target as a point mass and disregard its orientation. Consequently, these trackers often generate collision-free trajectories that awkwardly intrude into the target's frontal zone or fail to maintain a desired viewing angle (e.g., a continuous side-profile shot).To address this, we propose a pose-aware trajectory tracking framework that incorporates the target's orientation into both path finding and trajectory optimization. First, we design an anisotropic semantic-aware front-end that constructs a continuous potential field to penalize the target's frontal region, naturally inducing detour behaviors and providing a semantically valid initial path. Second, we extend the MINCO trajectory optimization framework to create a relative view-constrained back-end, utilizing the analytical gradient of the relative bearing constraint to ensure precise pose maintenance, safety, and smoothness. Simulation results demonstrate that, compared to existing baselines, our method significantly reduces frontal intrusions and improves relative view consistency in obstacle-dense environments.