Automated Detection of Korean Melons Based on YOLOv10 for Use in Quadruped Robot

The production of Korean melon, known as Chamoe, in South Korea faces significant challenges due to the decline of the rural labor force and the high labor demands of manual greenhouse harvesting. This study provides a critical step toward developing an automated Korean melon harvesting system using a quadruped robotic platform. This study focuses on the evaluation of a deep learning model, YOLOv10, for accurate melon detection in Korean greenhouse environments, which is a key component of the proposed automated harvesting system. We used YOLOv10, a state-of-the-art object detection algorithm, and the model was trained and tested on a custom dataset of Korean melon images. An NVIDIA RTX 4090 GPU and an Intel Core i9-13900K (3.0-5.8 GHz) CPU running Ubuntu 22.04 were used for the performance evaluation. The YOLOv10-based model showed promising results, achieving a mean average precision (mAP) of 0.927 at an intersection over union (IoU) threshold of 0.5. This model is suitable for use in robotic systems where resources are limited due to its compact size of only 5.7 MB. The model used only 11 % of the GPU capacity for setup and 2–3 % for inference. This demonstrates its efficiency and potential for real-time applications.