Authors

Abstract

To solve the problems of low precision in locating stem picking points and difficulty in recognizing occluded strawberry during the operation of strawberry picking robots, this paper proposed an improved YOLOv8-pose method for strawberry fruits identification and key points detection at the red ripe stage. Based on the YOLOv8-pose human posture estimation model, three categories (strawberry, stem, and picking points) were annotated. The acquired images were divided into training, validation, and test sets in an 8:1:1 ratio. In order to improve the feature extraction ability of the model for small targets, shuffle attention (SA) mechanism was added into the backbone network of YOLOv8-pose. Additionally, a comparative analysis was conducted to assess the impact of six attention mechanisms of CBAM (Convolutional block attention module), SimAM (Simple attention module), GAM (Global attention module), EMA (Efficient multi-scale attention), SK (Selective kernel attention), and SA on the detection results. Experimental results show that the proposed method can quickly and accurately detect strawberry fruits and key points for picking. The Precision (P), Recall (R), and mean average precision (mAP)50 values for both bounding boxes and key points based on SA mechanism were 99.7%, 100.0%, and 99.5% respectively, which were superior to the other attention mechanisms. Compared with YOLOv5-pose and YOLOv8-pose models, the improved model had the best P, R, and mAP50 values, and its memory usage was 6.4MB, which was also optimal. The improved method can provide crucial technical support for precise robotic strawberry picking.

Abstract in English

为解决草莓采摘机器人作业中果梗采摘点定位精度低和遮挡草莓识别困难等问题，本文以红熟期草莓为研究对象，提出一种改进的YOLOv8-pose草莓果实识别及采摘关键点检测方法。以YOLOv8-pose人体姿态估计模型为基础，标注了草莓、果梗、采摘点3个类别。将采集的图像按8:1:1的比例划分成训练集、验证集和测试集。为了提高模型对小目标的特征提取能力，在YOLOv8-pose的骨干网络中添加了shuffle注意力机制，并对比分析了CBAM（Convolutional block attention module），SimAM（Simple attention module），GAM（Global attention module），EMA（Efficient multi-scale attention），SK（Selective kernel attention）以及SA六种注意力机制对检测结果的影响。实验结果表明，本文提出的方法可以对草莓果实及采摘关键点进行快速准确检测，基于SA注意力机制的边界框和关键点的检测精确率（Precison，P），召回率（Recall，R），平均精度均值（Mean average precision，mAP）mAP50均为99.7%，100.0%和99.5%，均优于其他注意力机制；与YOLOv5-pose、YOLOv8-pose模型相比，改进模型的P，R，mAP50值都是最优的，模型的内存占用量为6.4MB，也是最优的，改模型可以为机器人精准采摘提供重要的技术支持。