

















摘" 要: 道路交通標志識別是自動駕駛、車聯網的重要組成部分,為進一步提高交通標志檢測的精度和速度,提出一種基于YOLOv8s改進的YOLOv8?TS道路交通標志檢測網絡。首先,對YOLOv8s進行了整體的輕量化設計,并設計了Conv?G7S和CSP?G7S模塊,減少了網絡的參數量;其次,設計了CSP?SwinTransformer模塊,強化了模型利用窗口內的特征信息進行上下文感知和建模的能力;然后,在頸部網絡融合了卷積注意力機制(CBAM),強化了模型對不同通道、空間權重信息的學習;最后,對損失函數進行了改進,提升了邊界框回歸性能。實驗結果表明,在中國道路交通標志TT100K數據集上,精確率(Precision)、平均精度(mAP@0.5)分別提高了6.9%、3.7%,而改進后模型的參數量下降了75.4%,模型的大小僅為5.8 MB,平均精度(mAP@0.5)達到96.5%,檢測速度由126.58 f/s提升至136.99 f/s。
關鍵詞: 交通標志檢測; YOLOv8?TS; 輕量化; 注意力機制; Conv?G7S; WIoU
中圖分類號: TN911.73?34; TP391.41" " " " " " " nbsp; "文獻標識碼: A" " " " " " " " "文章編號: 1004?373X(2025)01?0179?08
YOLOv8?TS traffic sign detection network integrating attention mechanism
HUANG Zhiyuan1, FANG Qiu1, 2, GUO Xinghao1
(1. School of Mechanical and Automotive Engineering, Xiamen University of Technology, Xiamen 361024, China;
2. School of Aerospace Engineering, Xiamen University, Xiamen 361005, China)
Abstract: Road traffic sign recognition is an important part of automatic driving and Internet of Vehicles (IoV). In view of this, the paper proposes an improved YOLOv8?TS road traffic sign detection network based on YOLOv8s to further improve the accuracy and speed of traffic sign detection. The overall lightweight design of the YOLOv8s is carried out. The Conv?G7S and CSP?G7S modules are designed to reduce the number of network parameters. The CSP?SwinTransformer block is designed to enhance the ability of the model to use the feature information in the window for context awareness and modeling. Then, CBAM (convolutional block attention module) is integrated in the neck network to strengthen the learning of different channels and spatial weight information. The loss function is improved to improve the performance of boundary box regression. The experimental results show that on the TT100K data set of Chinese road traffic signs, the precision and the mAP@0.5 are improved by 6.9% and 3.7%, respectively, while the parameters of the improved model decreases by 75.4%, its size is only 5.8 MB, its mAP@0.5 reaches 96.5%, and its detection speed is increased from 126.58 f/s to 136.99 f/s.
Keywords: traffic sign inspection; YOLOv8?TS; lightweight; attention mechanism; Conv?G7S; WIoU
0" 引" 言
交通標志檢測技術的研究與應用,不僅提升了無人駕駛汽車的實時感知能力,而且對于提醒駕駛員規范駕駛以減少交通事故也有重要意義。近些年來,涌現了許多與交通標志檢測相關的研究,主要分為傳統的特征提取方法[1]和深度學習方法,在傳統的方法中,研究人員通過分析交通標志的形狀、顏色等特征[2?4],手動設計對應的特征提取器來識別交通標志。近年來,隨著ResNet[5]、SSD[6]、Faster?RCNN、YOLO等模型的不斷涌現,基于深度學習的目標檢測技術發展迅速。
在交通標志檢測領域,文獻[7]提出了一種多尺度特征融合與極限學習機結合的交通標志識別方法,該方法融合了3個網絡模型來構建多尺度神經網絡,把提取的多尺度特征送入極限學習機,來實現交通標志的識別。……