Abstract
In the context of modern agricultural intelligence, smart picking is becoming a crucial method for enhancing production efficiency. This paper proposes a tomato detection and localization system that integrates the YOLOv5 deep learning algorithm with the SGBM algorithm to improve detection accuracy and three-dimensional localization of tomatoes in complex environments. In a greenhouse setting, 640 tomato images were collected and categorized into three classes: unobstructed, leaf-covered, and branch-covered.Data augmentation techniques, including rotation, translation, and CutMix, were applied to the collected images, and the YOLOv5 model was trained using a warmup strategy.Through a comparative analysis of different object detection algorithms on the tomato dataset, the feasibility of using the YOLOv5 deep learning algorithm for tomato detection was validated. Stereo matching was used to obtain depth information from images, which was combined with the object detection algorithm to achieve both detection and localization of tomatoes. Experimental results show that the YOLOv5 algorithm achieved a detection accuracy of 94.1%, with a distance measurement error of approximately 3-5 mm.