Robotic Attachment System for Hospital Bed and Trolley Transportation: A Hybrid Computer Vision Approach for Autonomous Hospital Logistics

Abstract

Autonomous mobile robots (AMRs) in hospital logistics require precise visual alignment for trolley docking. This thesis presents a hybrid computer vision approach that combines deep learning with geometric edge-based processing to detect both large trolley handles and small mounting holes in real time. The central challenge is multi-scale detection: handles occupy approximately 35% of the frame, while holes occupy approximately 0.14%. A YOLOv8n model trained on 39 annotated images (32 for training and 7 for validation, without a separate held-out test set) achieved reliable handle localization, but practical hole detection using machine learning alone remained unstable. To address this limitation, YOLO-based handle detection was used to define a region of interest, and Canny edge detection with contour-circularity filtering was used for hole localization. Subsystem-level evaluation showed a 96% hole-detection success rate in ROI-constrained tests and approximately 19–20 FPS processing on CPU hardware, with centimeter-level alignment estimation in controlled calibration cases. Due to the thesis timeframe, full physical end-to-end docking integration (perception, motion execution, and mechanical engagement in one autonomous sequence) was not completed; instead, perception and hardware subsystems were validated separately, and full integrated docking trials are planned as future teamwork. Results indicate that combining context-aware machine learning for larger objects with geometry-driven methods for small features is more robust than purely machine-learning-based or purely classical detection pipelines for this task.