PDF下载

[1]孙卉梅　刘路　王德刚　王太勇.面向高后果区工程车辆视觉检测的YOLO-MMCE算法[J].油气储运,2024,43(09):1-10.
　SUN Huimei,LIU Lu,WANG Degang,et al.An Improved YOLOv5 Algorithm for Visual Detection of Engineering Vehicles in High Consequence Areas[J].Oil & Gas Storage and Transportation,2024,43(09):1-10.

点击复制

面向高后果区工程车辆视觉检测的YOLO-MMCE算法

《油气储运》[ISSN:1000-8241/CN:13-1093/TE] 卷: 43 期数: 2024年09期页码: 1-10 栏目: 出版日期: 2024-09-25

Title:: An Improved YOLOv5 Algorithm for Visual Detection of Engineering Vehicles in High Consequence Areas

作者:: 孙卉梅　刘路　王德刚　王太勇

Author(s):: SUN Huimei; LIU Lu; WANG Degang; WANG Taiyong

关键词:: YOLOv5; 工程车辆; 数据增强; 坐标注意力机制; 损失函数

Keywords:: yolov5; engineering vehicles; data enhancement; ; coordinate attention mechanism; loss function

分类号:: TE832

文献标志码:: A

摘要:: 【目的】高后果区大型施工现场的工程车辆对埋地管道造成严重的安全隐患。针对当前常用方法在工程车辆重叠目标检测和日照变化场景下的目标检测方面存在漏检率高、检测精度低的问题，以挖掘机、装载机、压路机和重型货车4类常见工程车辆作为识别对象，提出了一种基于改进YOLOv5的工程车辆目标检测方法——YOLO-MMCE。【方法】采用Mosaic+Mixup结合的数据增强方式，增强对不同场景的适应能力，提高模型在实际复杂环境和模糊情况下的鲁棒性和泛化性。针对目标重叠和光照变化导致的特征不明显问题，在YOLOv5网络模型中引入坐标注意力机制（Coordinate Attention，CA），增强网络模型的特征提取能力；为了提升预测边框回归精度，引入了高效率交并比（Efficient Intersection over Union，EIOU）函数，计算了预测框和真实框的宽高差异值取代纵横比，进一步提高算法检测精度。【结果】以兰郑长成品油管道高后果区监控摄像机获取的施工现场照片为数据集，对YOLO-MMCE算法进行验证。实验结果表明，对YOLOv5算法3个方面的改进均能提高其在实际工况下工程车辆目标检测的精度，总体平均精度均值（Means Average Precision，mAP）达到84.8%，比原始YOLOv5算法提高了6.9%。对挖掘机、装载机、压路机和重型货车的目标检测mAP分别提高了4.4%、7.5%、9.5%、6.0%。【结论】YOLO-MMCE算法有效解决了重叠目标检测和日照变化场景下的工程车辆目标检测问题，具备实际应用价值。

Abstract:: [Objective]Engineering vehicles in large-scale construction sites in high-consequence areas pose serious long-term safety hazards to buried pipelines. Aiming at the problems of high leakage rate and low detection accuracy of current commonly used methods in overlapping target detection of engineering vehicles and target detection in sunlight change scenarios, we propose a target detection method for engineering vehicles based on improved YOLOv5 by taking four types of common engineering vehicles, including excavators, loaders, rollers and heavy trucks, as the recognition objects - -YOLO-MMCE. [Methods]The data enhancement method combining Mosaic+Mixup is adopted to enhance the adaptability to different scenes and improve the robustness and generalization of the model to the actual complex environment and fuzzy situations. To address the problem of feature inconspicuousness due to target overlapping and illumination changes, the Coordinate Attention (CA) mechanism is introduced into the YOLOv5 network model to enhance the feature extraction capability of the network model; in order to improve the accuracy of predicted border regression, the Efficient Intersection over Union (EIOU) function is introduced to enhance the regression accuracy of predicted borders, and the width-to-height difference between predicted and real borders is calculated to replace the aspect ratio to further improve the detection accuracy of the algorithm. [Results] The YOLO-MMCE algorithm is validated using the construction site photos obtained from the monitoring camera in the high consequence area of Lanzhou-Zhengzhou-Changsha oil pipeline as the dataset. The experimental results show that the improvement of all three aspects of the YOLOv5 algorithm can improve its accuracy of target detection of construction vehicles under actual working conditions, and the overall Means Average Precision (mAP) reaches 84.8%, which is 6.9% higher than the original YOLOv5 algorithm. The target detection mAP for excavators, loaders, rollers and heavy trucks were improved by 4.4%, 7.5%, 9.5% and 6.0%, respectively. [Conclusion] The YOLO-MMCE algorithm effectively solves the problems of overlapping target detection and target detection of engineering vehicles under the scenario of sunlight change, and has practical application value.

相似文献/References:

[1]孙卉梅,刘路,王德刚,等.面向高后果区工程车辆视觉检测的YOLO-MMCE算法[J].油气储运,2024,43(09):1031.[doi:10.6047/j.issn.1000-8241.2024.09.008]
　SUN Huimei,LIU Lu,WANG Degang,et al.YOLO-MMCE algorithm for visual detection of engineering vehicles in high-consequence area[J].Oil & Gas Storage and Transportation,2024,43(09):1031.[doi:10.6047/j.issn.1000-8241.2024.09.008]

更新日期/Last Update: 2024-06-26