基于改进SOLOv2算法的自主拾球机器人实例分割研究  

Research on Instance Segmentation of Autonomous Ball-Picking Robots Based on the Improved SOLOv2 Algorithm

作　　者：许广胜 官源林 杨熙鑫 董建林 XU Guangsheng;GUAN Yuanlin;YANG Xixin;DONG Jianlin(College of Mechanical and Automotive Engineering,Qingdao University of Technology,Qingdao 266525,China;College of Computer Science and Technology,Qingdao University,Qingdao 266071,China)

机构地区：[1]青岛理工大学机械与汽车工程学院,山东青岛266525 [2]青岛大学计算机科学技术学院,山东青岛266071

出　　处：《青岛大学学报(工程技术版)》2025年第1期23-31,共9页Journal of Qingdao University(Engineering & Technology Edition)

基　　金：国家自然科学基金资助项目(52375348);山东省高等学校科技计划资助项目(J17KA047;J17KA055);青岛市自然科学基金资助项目(23-2-1-216-zyyd-jch)。

摘　　要：针对拾球机器人原有SOLOv2算法中残差网络存在的训练耗时严重、梯度消失和过拟合等问题,提出基于随机深度与AdamW优化SOLOv2的图像实例分割方法。基于随机深度算法的ResNet网络,在模型训练时剔除冗余的残差块,有效提高训练效率;采用AdamW优化算法自适应地调整改进SOLOv2网络中各参数的学习率,通过权重衰减优化网络的正则化项,防止系统过拟合,并以自定义COCO数据集的图像实例分割为例进行实验验证。实验结果表明,与经典SOLOv2算法相比,改进的SOLOv2算法实例分割精度达到了73.7%,提升了12.35%,有效缩短了模型训练时间,提升了模型识别的鲁棒性,实现机器人自动取球与放球功能。To address issues such as significant training time,gradient vanishing,and overfitting in the original SOLOv2 algorithm of ball-picking robots,this study proposes an image instance segmentation method for optimizing SOLOv2 based on stochastic depth and the AdamW optimizer.The ResNet network with stochastic depth removes redundant residual blocks during model training,effectively improving training efficiency.The AdamW optimization algorithm adaptively adjusts the learning rates of various parameters in the improved SOLOv2 network and enhances network regularization through weight decay,preventing system overfitting.Experimental validation is conducted using a custom COCO dataset for image instance segmentation.The results show that,compared to the classic SOLOv2 algorithm,the improved SOLOv2 algorithm achieves an instance segmentation accuracy of 73.7%,an improvement of 12.35%.It also significantly reduces model training time,enhances the robustness of model recognition,and enables the robot to perform automatic ball-picking and placement functions effectively.

关 键 词：SOLOv2 随机深度 AdamW ResNet 实例分割 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术] TP242.2[自动化与计算机技术—计算机科学与技术]