Development of a Drone Flight Simulation Framework for Autonomous Plant Counting Missions in Agriculture

This thesis presents a reproducible simulation framework for autonomous UAV-based plant counting built on Gazebo, PX4 SITL, and ROS 2. The framework combines a synthetic agricultural world with fixed ground truth (186 plants), autonomous lawnmower missions, mission-state-aware perception gating, and a detector-agnostic counting pipeline. Two candidate generators were integrated: an HSV-based baseline and a YOLOv8n detector trained on simulation-derived images. In both cases, detections were passed to shared downstream logic comprising centroid extraction, temporal association, track confirmation, stabilization, and optional world-space duplicate suppression. The framework was evaluated through controlled experiments that varied altitude, gating policy, tracking parameters, and duplicate-handling settings, with repeated runs for each configuration. The results show that plant-counting accuracy depends on the interaction between detector behavior and mission design rather than on detection quality alone. In the tested simulator setting, HSV provided the strongest performance for online mission-time plant counting, while YOLO improved substantially under global offline consolidation and became competitive when broader post-flight merging was allowed. The main contribution of the thesis is therefore not only a detector comparison, but an experimentally validated, mission-aware simulation framework for studying autonomous plant counting as an integrated autonomy-perception problem. The framework also provides a basis for future work on detector robustness, orthomosaic-based post-flight counting, and sim-to-real transfer.