基于改进YOLO v8-Pose的串番茄采收果柄二次剪切点识别与定位

Active Recognition and Localization of Secondary Cutting Point for Cherry Tomato in Intelligent Harvesting Based on Improved YOLO v8-Pose

摘要:

针对串番茄二次剪柄作业中存在的剪切点易被端部果实遮挡、定位精度不高和剪切易损伤果实的问题,本文提出基于改进YOLO v8-Pose的解决方法：利用主动识别的思想,依据4分类＋空间平面法向量的方法旋转果串以确定最佳观测角度;利用关键点信息定位二次剪切点。在网络结构方面,引入PSA部分自注意力机制以增强全局上下文建模能力;设计具有多尺度卷积核的Dyn_GSConv模块,并基于Slim-Neck结构构建Dyn_VoVGSCSP模块以替代原颈部的C2f结构,在降低模型复杂度的同时兼顾局部细节与全局特征;最后采用VCM模块替换特征提取层的标准卷积,进一步精简网络。试验结果表明,改进后网络在类别检测与关键点识别的mAP50均达到97.4%,较原模型提升1.4个百分点;精确率分别为93.3%和90.9%,提升5.3、5.0个百分点。同时,浮点运算量降低9×108,参数量减少1.9×105。在最佳观测视角选取方面,根据网络类别输出将随机姿态的果串调整至正面,并结合深度信息构建串番茄空间平面法向量,当法向量与相机yz面差角小于10°,即确定最佳观测视角,为关键点识别和二次剪切操作提供清晰观测条件。本研究为实现串番茄自动化、高精度、低损伤的二次剪柄作业提供了有效的视觉感知解决方案。

Abstract:

Aiming to address issues in the secondary stem-cutting operation of cherry tomato,such as occlusion of cutting points by terminal fruits,low positioning accuracy,and risk of fruit damage,a solution based on an improved YOLO v8-Pose was proposed. By introducing an active recognition strategy,the tomato was rotated to determine the optimal viewing angle according to a four-category classification and spatial plane normal vector method,while key points were used to locate secondary cutting positions. In terms of network architecture,the PSA partial self-attention mechanism was integrated to enhance global context modeling. A Dyn_GSConv module with multi-scale convolutional kernels was designed,and a Dyn_VoVGSCSP module was constructed based on a Slim-Neck structure to replace the original C2f neck,reducing model complexity while preserving local details and global features. Meanwhile,a VCM module was adopted to replace standard convolution in the feature extraction layer,further lightweighting the network. Experiments showed that the improved network achieved an mAP50 of 97.4% in both detection and keypoint recognition,an improvement of 1.4 percentage points over the original model. The precision rates reached 93.3% and 90.9%,increasing by 5.3 percentage points and 5.0 percentage points,respectively. Furthermore,GFLOPs were reduced by 9×108,and the number of parameters was decreased by 1.9×105. For optimal viewing angle selection,the network's category output was used to adjust randomly posed tomatoes to a front view,and depth information was employed to construct spatial plane normals of fruit clusters. When the angle between the normal vector and the camera's yz-plane was less than 10°,the optimal viewing angle was determined,providing a clear observation view for keypoint recognition and secondary cutting operations. The research result can provide an effective visual perception solution for achieving automated,high-precision,and low-damage secondary stem-cutting in cherry tomato production.