Authors

Abstract

Target recognition and localization of picking robots are very important in orchard environments. This paper analyzed the You Only Look Once version 8 normal (YOLOv8n) method in deep learning for citrus picking robots, introduced FasterNet, efficient multi-scale attention (EMA) mechanism, and Wise Intersection over Union (WIoU) loss function, developed an improved YOLOv8n method, and collected a citrus fruit dataset to analyze the recognition and positioning effects of the proposed methods. The selected FasterNet, EMA, and WIoU were all superior to the other types compared, and the improved YOLOv8n method achieved an accuracy of 0.893, a recall rate of 0.894, and a mean average percentage (mAP) of 0.895 for citrus recognition, outperforming other recognition methods. In terms of citrus localization, the average errors on X, Y, and Z axes were 3.56 mm, 3.54 mm, and 4.74 mm, respectively. The findings demonstrate the reliability of the proposed model, and it can be applied to actual citrus picking robots.

Abstract in English

在果园环境下，采摘机器人的目标识别与定位十分重要。本文针对柑橘采摘机器人，分析了深度学习中的YOLOv8n方法，引入FasterNet、EMA注意力机制以及WIoU损失函数，设计了一种改进YOLOv8n，并采集柑橘果实数据集，对所提方法的识别和定位效果进行了分析。结果发现，所选择的FasterNet、EMA以及WIoU均优于所比较的其他类型，改进YOLOv8n在柑橘识别上获得了0.893的准确率、0.894的召回率和0.895的mAP，优于其他识别方法。在柑橘定位上，在X、Y、Z上的平均误差分别为3.56mm、3.54mm和4.74mm。结果证明了所提方法的可靠性，可以在实际的柑橘采摘机器人中进行应用。