基于YOLOv5的机耕船双目视觉障碍感知研究  

Research on binocular visual impairment perception of a cultivator boat based on YOLOv5

作　　者：陈泉淦 陈新元[1,2] 曾镛[1,2] 程志文 Chen Quangan;Chen Xinyuan;Zeng Yong;Cheng Zhiwen(School of Mechanical Automation,Wuhan University of Science and Technology,Wuhan,430081,China;Key Laboratory of Metallurgical Equipment and Control of Ministry of Education,Wuhan University of Science and Technology,Wuhan,430081,China;Wuhan Huayou Tianyu Technology Co.,Ltd.,Wuhan,430210,China)

机构地区：[1]武汉科技大学机械自动化学院,武汉市430081 [2]武汉科技大学冶金装备及其控制教育部重点实验室,武汉市430081 [3]华友天宇科技(武汉)股份有限公司,武汉市430210

出　　处：《中国农机化学报》2024年第7期261-268,共8页Journal of Chinese Agricultural Mechanization

基　　金：湖北省农机装备补短板核心技术应用攻关项目(HBSNYT202204);武汉市企业技术创新项目(2020020602012133)。

摘　　要：为满足机耕船自动驾驶功能,设计一套YOLOv5融合SGBM算法的机器视觉障碍感知系统。首先,以人、机耕船和农具为对象拍摄和收集图片得到水田障碍数据集,将图像输入YOLOv5网络模型迭代训练得到最优权重,随后将最优权重用于测试,并且与YOLOv4和Faster R-CNN网络进行比较;将双目相机拍摄的左右图像输入YOLOv5模型中进行检测,将输出的目标障碍检测框信息经校正变换后用SGBM算法进行视差计算,完成对目标障碍的识别和深度估计。结果表明,YOLOv5的平均精度均值稳定在87.25%比YOLOv4高1.55%,比Faster R-CNN高4.04%,单幅图像检测时间为0.017 s比YOLOv4快0.081 s,比Faster R-CNN快0.182 s且模型大小仅为13.7 MB比YOLOv4小236.4 MB;在检测机耕船、人和农具时,YOLOv5网络模型的置信度分别为0.91、0.99、0.95。YOLOv5+SGBM的深度估计在2 m内,精度达到98.1%。基于YOLOv5和SGBM的水田深度估计,能满足带旋耕无人驾驶作业的机耕船实际需求。In order to satisfy the automatic driving function of the boat tractor,this paper designed a set of YOLOv5 integrated SGBM algorithm machine vision obstacle perception system.Firstly,people,machine-tiller and farm tools were taken as objects to shoot and collect images to get paddy field obstacle data set.The images were input into the YOLOv5 network model for iterative training to get the optimal weight.Later,the most weight was used for testing and compared with YOLOv4 and Faster R-CNN networks.The left and right images taken by the binocular camera were input into the YOLOv5 model for detection.After correcting and transforming the output information of the target obstacle detection box,the SGBM algorithm was used for parallax calculation to complete the target obstacle recognition and depth estimation.The results show that the average accuracy of YOLOv5 is stable at 87.25%,1.55%higher than that of YOLOv4,4.04%higher than that of Faster R-CNN,and the detection time of a single image is 0.017 s,0.081 s faster than that of YOLOv4.It is 0.182 s faster than Faster R-CNN,and the model size is only 13.7 MB,236.4 MB smaller than YOLOv4.The confidence of the YOLOv5 network model is 0.91,0.99 and 0.95 respectively when detecting the boat tractor,man and farm tools.The depth estimation of YOLOv5+SGBM within 2 m,and the accuracy reaches 98.1%.The paddy field depth estimation based on YOLOv5 and SGBM can meet the actual requirements of unmanned boat tractor with rotary tillage.

关 键 词：机耕船 障碍感知 机器视觉 YOLOv5 深度估计 

分 类 号：S24[农业科学—农业电气化与自动化]