| 摘要: |
| 针对当前视觉黑烟车辆检测精度低、小目标难以检测的问题，提出改进YOLOv5s的黑烟车辆检测算法。首先，基于公开网络数据和真实道路拍摄图像构建黑烟车辆数据集，解决数据集受限问题。其次，改进网络模型，添加预测层，提高模型对小目标的检测性能，引入坐标注意力（Coordinate Attention，CA），增强模型的特征提取能力，进一步提高检测精度。最后，改进边界框回归损失函数为GIoU，提高边界框定位精度。实验结果表明，该改进模型能够有效地检测远距离小目标，改善漏报和虚警等问题。与原始YOLOv5s模型相比，该改进模型平均检测精度（mAP）提高3.1%，黑烟类别检测精度（AP）提高4.9%，在小目标场景中表现出较强的泛化能力。 |
| 关键词: 深度学习 YOLOv5s 黑烟车辆检测 小目标检测 大气环境监测 |
| DOI：DOI：10.3969/j.issn.1674-6732.2024.04.008 |
| 分类号:X831 |
| 基金项目:黑龙江省自然科学基金基项目（E2017001）；国家重点研发计划（2017YFC0803901-2） |
|
| Smoky Vehicle Detection Algorithm Based on Improved YOLOv5s Research |
|
MENG Yiqing1， WU Xiaodong1*， LI Bing2， GUAN Jiajun1
|
|
1.School of Civil Engineering and Transportation， Northeast Forestry University， Harbin， Heilongjiang 150040，China；
2.School of Mechanical and Electrical Engineering， Northeast Forestry University， Harbin， Heilongjiang 150040，China
|
| Abstract: |
| An improved YOLOv5s algorithm of smoky vehicle detection has been established to address the issue of low resolution and too small target to test. Firstly， the smoke vehicle dataset is constructed using public network data and real road photography to solve the problem of a limited dataset. Secondly， the introduction of Coordinate has further optimized the model. In addition， the paper introduces Coordinate and improves the regression loss function of the bounding box into GIOU to improve the location precision of the bounding box. Experiments show that the proposed model is effective in detecting small targets over a long distance， and it can solve the problem of false negative and false alarm. A comparison with the original YOLOv5s model reveals an increase of 3.1% in the average detection accuracy（mAP） and a 4.9% enhancement in the detection accuracy of the black smoke category（AP）， the model exhibits strong generalization ability in small target scenarios. |
| Key words: Deep learning YOLOv5s Smoke detection Small target detection Atmospheric environment monitoring |
