基于YOLOv8的车载红外目标检测改进算法研究  

Research on Vehicle Infrared Target Detection Improved Algorithm Based on YOLOv8

作　　者：侯军 杨洁[1] 邵凯青 HOU Jun;YANG Jie;SHAO Kaiqing(College of Machinery and Transportation,Southwest Forestry University,Kunming,Yunnan 650224,China)

机构地区：[1]西南林业大学机械与交通学院,云南昆明650224

出　　处：《计量学报》2025年第2期167-176,共10页Acta Metrologica Sinica

基　　金：云南省教育厅科学研究基金(0111723084,2024Y608)。

摘　　要：针对车载红外图像检测中的目标相互遮挡和小尺度目标漏检问题,提出一种基于YOLOv8的车载红外目标检测改进算法(VITD-YOLO)。首先,在Neck网络中增加大尺寸特征网络预测层(S-layer),增强网络对于小目标的检测精度;其次,在Backbone网络中设计C2F-DA模块,利用offset轻量化结构增强模型对目标的局部特征感知能力,并结合3种不同尺度自注意力设计了动态卷积头检测模组(Dy-head),提高被遮挡和密集目标的定位和分类精度;最后,采用Focal-SIoU作为网络的损失函数,解决训练样本中行人车辆目标类别不均衡问题,并提高网络训练和推理能力。将该算法在FLIR红外数据集上测试,实验结果表明:VITD-YOLO具有良好的检测效果和鲁棒性,对小尺度目标检测精度更高;该算法的平均精度达到91.2%,比原算法提高了2.5%,召回率达到83.4%,比原算法提高3.2%。Aiming at the problem of mutual occlusion of targets and missed detection of small-scale targets in vehiclemounted infrared image detection,an vehicle-mounted infrared target detection improved algorithm(VITD-YOLO)based on YOLOv8 is proposed.Firstly,a large-size feature network prediction layer(S-layer),is added to the Neck network to enhance the detection accuracy of the network for small objects.Secondly,the C2F-DA module is designed in Backbone network,and the offset lightweight structure is used to enhance the perception ability of the model to the local features of the target.A dynamic convolutional head(Dy-head)detection module is designed by combining three different scales of self-attention to improve the location and classification accuracy of occluded and dense targets.Finally,Focal-SIoU is used as the loss function of the network to solve the problem of class imbalance of pedestrian and vehicle targets in the training samples and improve the training and reasoning ability of the network.The algorithm is tested on FLIR infrared data set.Experimental results show that VITD-YOLO has good detection effect and robustness and higher accuracy for small-scale target detection.The average precision of the algorithm reaches 91.2%,which is 2.5%higher than that of the original algorithm,and the recall rate reaches 83.4%,which is 3.2%higher than that of the original algorithm.

关 键 词：机器视觉 车载红外目标检测算法 YOLOv8 辅助驾驶 图像识别 C2F-DA Focal-SioU 

分 类 号：TB96[机械工程—光学工程] TP391.41[一般工业技术—计量学]