摘要
交通标志检测是自动驾驶、智能交通系统以及道路安全监控等的关键任务之一。针对交通标志检测存在小目标数量多、精度低、传统模型体积大、不宜部署等问题,基于YOLOv8n网络模型提出了一种新的交通标志检测算法—Faster-YOLOv8。该模型在Neck部分采用C2f-Faster模块(C2f和FasterNet的高效融合)来优化YOLOv8n网络结构,降低模型参数量及模型大小;引入EMA注意力机制,并应用于模型主干网络,实现了更好的多尺度感知和空间感知,增强了模型的特征提取能力;通过添加小目标检测层,有效地结合了不同尺度特征信息,保留更多的细节信息,从而提高了对小目标的检测能力;采用SioU作为边界损失函数提高检测精度。研究结果表明:改进的Faster-YOLOv8在中国交通标志检测数据集TT100K中的检测精度(P_(recision))、召回率(R_(ecall))、平均精度均值(m_(AP@0.5))分别达到了79.8%、69.3%、77.8%,相比YOLOv8n模型提升了1.1%、2.8%、2.9%,参数量及模型大小减少了23.59%、19.16%,在检测准确性和模型轻量化方面都取得了显著改进,与现有方法相比,具有更强的实际应用价值。
Traffic sign detection is one of the crucial tasks in applications such as autonomous driving,intelligent traffic systems,and road safety monitoring.In response to the problems such as a large number of small target objects,low accuracy,large volume of tradition models and unsuitable deployment in traffic sign detection,a novel traffic sign detection algorithm based on YOLOv8n network model,that is Faster-YOLOv8,was proposed.In the Neck section,the proposed model optimized the network structure of YOLOv8n by employing the C2f-Faster module(efficient fusion of C2f and FasterNet),which reduced the number of model parameters and model size.Furthermore,EMA attention mechanism was introduced to the backbone network of the model to realize better multi-scale and spatial perception,which improved feature extraction of the model.Additionally,a small target detection layer was added to effectively combine feature information from different scales and preserve more detailed information,thereby enhancing the detection ability of small objects.Finally,SIoU was utilized as the boundary loss function to improve detection accuracy.The research results demonstrate that the improved Faster-YOLOv8 achieves detection accuracy(P_(recision)),recall rate(R_(ecall)),and mean average precision(m_(AP@0.5))of 79.8%,69.3%,and 77.8%,respectively,in the Chinese traffic sign detection dataset TT100K.Compared to the YOLOv8n model,it exhibits an improvement of 1.1%,2.8%,and 2.9%in these metrics,while reducing model parameters and size by 23.59%and 19.16%,respectively.The proposed model significantly enhances both detection accuracy and model lightweighting,demonstrating practical utility superior to the existing methods.
作者
高良鹏
赵博文
简文良
GAO Liangpeng;ZHAO Bowen;JIAN Wenliang(College of Transportation,Fujian University of Technology,Fuzhou 350118,Fujian,China)
出处
《重庆交通大学学报（自然科学版）》
CAS
CSCD
北大核心
2024年第8期114-123,共10页
Journal of Chongqing Jiaotong University(Natural Science)
基金
福建省自然科学基金项目(2020J05194,2021J05226,2023J01946)。
作者简介
第一作者:高良鹏(1988-),男,福建漳州人,副教授,博士,主要从事交通规划与管理方面的研究。E-mail:liangpenggao.acad@gmail.com;通信作者:简文良(1991-),男,福建漳州人,副教授,博士,主要从事交通运输经济方面的研究。E-mail:wenliang_jian@fjut.edu.cn。
