DAI J, LI L X, ZHAO J W, et al.Research on garbage grasping object detection algorithm based on improved YOLOv5s[J].Journal of Henan Polytechnic University(Natural Science) ,2026,45(3):49-58.

基于改进YOLOv5s的垃圾抓取目标检测算法研究

代军1, 李理想1, 赵俊伟1, 袁兴起1, 刘钢1, 程晓琦2

1.河南理工大学 机械与动力工程学院，河南 焦作 454003；2.佛山大学 机电工程与自动化学院，广东 佛山 528225

摘要:目的 垃圾分类智能化是未来发展的趋势，为提高垃圾分类和抓取的精确度和速度，降低人工成本，提出一种基于改进YOLOv5s的垃圾检测方法。  方法 首先，YOLOv5s骨干网络采用参数更少、计算更快的MobileNet取代之前的CSPDarkNet53，并在网络的检测头部引入CBAM注意力机制；其次，构建垃圾分类数据集并进行数据增强，在该数据集上训练Improved-YOLOv5s模型，并与原始YOLOv5s模型进行实验对比；最后，将改进后的网络模型移植到垃圾抓取机器人平台中进行抓取测试实验，该平台包括控制系统、视觉系统、抓取系统和移动基座。  结果 实验结果显示，与原始YOLOv5s模型相比，改进后的Improved-YOLOv5s模型参数量为3.63×106，仅为原始模型的51.63%；训练速度提高了60%，为5 min/次；检测速度提高了20%，为33 fps；检测精度mAP为0.44。抓取实验结果表明，机器人对形状规则、重量较轻的垃圾有较好的抓取表现，其中废纸团抓取成功率为95%，塑料抓取成功率为85%。  结论 所提目标检测算法能够很好平衡检测精度和算法复杂度，比原始YOLOv5s更为轻量化，适合部署到移动端和嵌入式平台，提高了垃圾抓取机器人的实时性和成功率，对垃圾分类和智能环卫的发展具有一定的实用价值。

关键词:目标检测;垃圾抓取;轻量化神经网络;垃圾分类;智能化

doi:10.16186/j.cnki.1673-9787.2023110027

基金项目:国家自然科学基金资助项目（62201151）；河南省科技攻关项目（232102221028）；河南省高等学校重点科研项目（22A460020）；河南理工大学博士基金资助项目（B2016-22）

收稿日期:2025/02/13

修回日期:2025/06/31

出版日期:2026/04/28

Research on garbage grasping object detection algorithm based on improved YOLOv5s

Dai Jun1, Li Lixiang1, Zhao Junwei1, Yuan Xingqi1, Liu Gang1, Cheng Xiaoqi2

1.School of Mechanical and Power Engineering， Henan Polytechnic University， Jiaozuo  454003， Henan， China；2.School of Mechatronic Engineering and Automation， Foshan University， Foshan  528225， Guangdong， China

Abstract: Objectives Intelligent waste classification is a future development trend. To improve the accuracy and speed of waste classification and grasping while reducing labor costs， a waste detection method based on an improved YOLOv5s is proposed.  Methods First， the backbone network of YOLOv5s is replaced with MobileNet， which has fewer parameters and faster computation， instead of the original CSPDarkNet53. The CBAM attention mechanism is introduced into the detection head of the network. Second， a waste classification dataset is constructed and augmented. The improved YOLOv5s （Improved YOLOv5s） model is trained on this dataset and compared with the original YOLOv5s model through experiments. Finally， the improved network model is deployed on a waste grabbing robot platform （including control system， vision system， grasping system， and mobile base） for grasping tests.  Results Experimental results show that， compared with the original YOLOv5s model， the Improved YOLOv5s model has 3.63×10⁶ parameters， which is only 51.63% of the original model. The training speed is increased by 60% （5 min/epoch）， and the detection speed is increased by 20% （33 fps）. The mean average precision （mAP） is 0.44. Grasping experiments indicate that the robot performs well on regularly shaped， lightweight waste： the success rate for waste paper balls is 95%， and for plastic items it is 85%.  Conclusions The proposed object detection algorithm achieves a good balance between detection accuracy and algorithmic complexity. It is more lightweight than the original YOLOv5s， making it suitable for deployment on mobile and embedded platforms. It improves the real time performance and success rate of the waste grabbing robot， and provides practical value for the development of intelligent waste classification and sanitation.

Key words:object detection;garbage grasping;lightweight neural network;waste classification;intelligence