摘要
当前交通标志检测方法主要依赖单阶段深度学习算法构建的目标检测模型,存在检测精度低、模型通用性弱等问题。为解决这类问题,提出一种基于改进YOLOv8模型的交通标志检测方法。该方法通过引入基于注意尺度序列融合的机制,提升了神经网络对于多尺度信息的提取能力;通过增加小目标检测层,使得方法更适用于小目标检测;采用RT-DETR的检测头,通过解耦尺度内交互和跨尺度融合高效处理多尺度特征。此外,为了克服现有交通标志检测方法在弱泛化方面的局限性,提高包围盒回归的准确性和效率,采用一种全新的损失函数inner-mpdiou,有效提高了模型的训练效率和精度。基于清华-腾讯100K(TT100K)数据集的实验结果表明:在保证实时性的前提下,该方法平均精度高达84.0%,相较于目前国际主流YOLOv8模型,提高了7.1%,整体模型大小降低了12.9%,提升了低分辨小目标检测有效性。
Current methods for traffic sign detection primarily rely on single-stage deep learning algorithms to construct target detection models,which suffer from low detection accuracy and weak model generalizability.To address these issues,a traffic-sign detection method based on an improved the YOLOv8 model was proposed.This method introduces an attention-scale sequence fusion mechanism that enhances the ability of the neural network to extract multi-scale information.The addition of a small-object detection layer makes this method more suitable for small-object detection.Additionally,it adopts the RT-DETR detector head to process multi-scale features efficiently by decoupling the intra-scale interactions and cross-scale fusions.Moreover,to overcome the limitations of existing methods in terms of weak generalization and to improve the accuracy and efficiency of bounding box regression,a novel loss function,innermpdiou,is employed,which effectively improves the training efficiency and accuracy of the model.Experimental results on the TsinghuaTencent 100K(TT100K)dataset showed that under the premise of real-time performance,this method achieved an average accuracy of up to 84.0%.Compared with the current international mainstream YOLOv8 model,its average accuracy was improved by 7.1%,and the overall model size was reduced by 12.9%,thereby enhancing the effectiveness of low-resolution small object detection.
作者
余荣威
张逸轩
曹书明
王丽娜
YU Rongwei;ZHANG Yixuan;CAO Shuming;WANG Lina(School of Cyber Science and Engineering,Wuhan University,Wuhan 430072,Hubei,China;Key Laboratory of Aerospace Information Security and Trustworthy Computing,Ministry of Education,Wuhan 430072,Hubei,China)
出处
《武汉大学学报(理学版)》
北大核心
2025年第4期453-462,共10页
Journal of Wuhan University(Natural Science Edition)
基金
国家重点研发计划(2022YFB4500800)
国家自然科学基金(42071431)。
