农业机械学报2025，Vol.56Issue(10)：558-566,9.DOI:10.6041/j.issn.1000-1298.2025.10.050
基于YOLO v5-MNv4S与RANSAC-GN的柑橘姿态估计方法
Citrus Pose Estimation Based on YOLO v5-MNv4S and RANSAC-GN
摘要
Abstract
Aiming to address the challenge of estimating the variable pose of citrus in orchard environments and enhance the citrus harvesting success rate,a lightweight target detection network,YOLO v5-MNv4S,along with an improved random sample consensus-Gauss Newton(RANSAC-GN)point cloud processing algorithm was introduced.The citrus pose was defined based on actual citrus growth and the final harvesting process,constructing a real-time citrus pose estimation system.The YOLO v5s backbone network was optimized to be a lightweight feature extraction network(MNv4-Conv-S),which significantly reduced the training parameters so that the final output network weights were lightweight,reducing the amount of computation and improving the recognition efficiency.Additionally,the CA attention mechanism was incorporated,and the loss function was replaced with SIoU,addressing the weak feature extraction ability of the lightweight network.These improvements resulted in a lightweight YOLO v5-MNv4S network with superior detection capabilities compared with YOLO v5s.After the D435i image acquisition,YOLO v5-MNv4S was input to detect the target bounding box by using the pinhole model to output the citrus regional point cloud,segment the citrus surface point cloud,combined with the improved RANSAC-GN point cloud algorithm to fit accurate and stable citrus parameters,then fusion of the stem-end spatial coordinates,and ultimately the output of citrus spatial pose results to be harvested.Ablation experiments and network comparisons demonstrated that the lightweight YOLO v5-MNv4S achieved 93.1％accuracy,with only 14.7％of the parameters found in YOLO v5s.Compared with YOLO-Ghost,YOLO v7,YOLO v8,and other networks,it offered the best recognition accuracy with significantly reduced parameters.Experimental results for citrus localization and pose recognition showed that the citrus parameter fitting error using RANSAC-GN was(0.18,0.19,0.44)mm,and the pose estimation error was 2.56°.The pose estimation was accurate,and the estimation results of citrus pose in real orchard environments were consistent with real citrus.The research result can recognize citrus pose in orchard environments and provide technical support for structured citrus orchard mechanical harvesting equipment.关键词
柑橘采摘/目标检测/姿态估计/轻量化网络/点云处理/MobileNetV4Key words
citrus picking/object detection/pose estimation/lightweight network/point cloud processing/MobileNetV4分类
计算机与自动化引用本文复制引用
李丽,张官明,张云峰,梁继元,淳长品..基于YOLO v5-MNv4S与RANSAC-GN的柑橘姿态估计方法[J].农业机械学报,2025,56(10):558-566,9.基金项目
重庆市杰出青年科学基金项目(2022NSCQ-JQX0030)、宜宾市双城协议保障科研经费科技项目(XNDX2022020015)、中央高校基本科研业务费项目(Swu-XDJH202302)和重庆市研究生科研创新项目(CYB23125) （2022NSCQ-JQX0030）
