A Behavior Tree-Based Visual Reasoning Architecture for Enhanced UAM Object Recognition

With the rise of Urban Air Mobility, vision-based object recognition using ground-fixed observation systems is gaining significance as a key component for ensuring aerial safety. In urban environments, however, consistent and reliable recognition of UAM vehicles remains a challenge due to complex visual backgrounds, the small size of aircraft, and their irregular flight trajectories. To address these issues, this study proposes a visual inference framework based on Behavior Tree. Inspired by human visual cognition, the proposed recognition pipeline mimics the process in which a human observer first applies global attention to a scene and then focuses selectively on uncertain areas for further analysis. Following an initial recognition stage using YOLOv8, ROI are re-examined through SAHI-based slicing to enhance recognition of small or partially missed objects. This modular structure separates the processes of image input, primary recognition, ROI inference, and result storage, while the Behavior Tree framework enables clear execution flow and real-time state tracking through explicit node transitions. Experimental evaluations conducted in a ROS2-based environment demonstrate that the proposed architecture significantly outperforms a conventional single-stage YOLO system, particularly under conditions involving cluttered backgrounds and small-scale aerial targets. Additionally, per-frame state logging and automatic result saving enhance system analysis and debugging capabilities. This work validates Behavior Tree-based visual inference as an effective control framework for object recognition systems and suggests its potential for future integration with real-time autonomous flight and control systems.