基于Swin-Transformer的YOLOX交通标志检测  

YOLOX Traffic Sign Detection Based on Swin-Transformer

作　　者：嵇文 刘全金 黄崇文 杨瑞 黄汇磊 徐光豪 JI Wen;LIU Quanjin;HUANG Chongwen;YANG Rui;HUANG Huilei;XU Guanghao(School of Electronic Engineering and Intelligent Manufacturing,Anqing Normal University,Anqing 246011,China;College of Information and Electronic Engineering,Zhejiang University,Hangzhou 310027,China)

机构地区：[1]安庆师范大学电子工程与智能制造学院,安徽安庆246011 [2]浙江大学信息与电子工程学院,浙江杭州310027

出　　处：《无线电通信技术》2023年第3期547-555,共9页Radio Communications Technology

基　　金：国家重点研发计划(2021YFA1000502);国家自然青年基金(62101492);浙江省杰出青年基金(LR22F010002)。

摘　　要：交通标志检测是驾驶辅助系统和自动驾驶系统的关键因素之一。在交通标志检测过程中,交通标志距离不同导致目标尺度变化很大,远距离小尺度交通标志对基于卷积网络的目标检测器提出了巨大挑战。YOLOX-Swin算法将Swin-Transformer作为YOLOX的骨干网络以提取交通标志图像特征,通过移动窗口获取足够的全局上下文信息,并利用多头自注意力机制提取更多差异化特征;利用YOLOX自身的路径增强特征金字塔网络(Path Aggregation Feature Pyramid Network, PAFPN)提取、融合包括交通标志低层信息在内的多尺度特征信息,提升小目标交通标志检测精度。由于小目标交通标志在图像中所占像素较少,同时考虑到Transformer需要的训练样本多于卷积网络,在原本的复制粘贴法上进行改进,增加交通标志样本数量,以进一步提高交通标志检测精度。在TT100K数据集上的测试结果表明,所提目标检测方法较其他几种方法具有更高的交通标志检测精度,能满足交通标志检测准确性和实时性要求。Traffic sign detection is a crucial factor in driving assistance systems and autonomous driving systems.In the process of traffic sign detection,the varying distances of traffic signs cause significant challenges for object detectors based on convolutional neural networks,particularly for small-scale traffic signs at far distances.The YOLOX-Swin algorithm uses the Swin Transformer as the backbone network to extract image features of traffic signs.By using a sliding window approach,sufficient global contextual information is obtained,and a multi-head self-attention mechanism is utilized to extract more differentiated features.The YOLOX􀆳s Path Aggregation Feature Pyramid Network(PAFPN)is then used to extract and fuse multi-scale feature information,including low-level information of traffic signs,to improve the accuracy of small target traffic sign detection.Due to the limited number of pixels occupied by small target traffic signs in images and the need for more training samples for Transformers than convolutional networks,the original copy-paste method is improved by increasing the number of traffic sign samples to further enhance the accuracy of traffic sign detection.Test results on the Tsinghua-Tencent 100K(TT100K)dataset show that the proposed object detection method has higher accuracy in traffic sign detection than other methods,meeting the requirements for accuracy and real-time detection of traffic signs.

关 键 词：深度学习 YOLOX Swin-Transformer 小目标检测 复制粘贴法 

分 类 号：TP391.4[自动化与计算机技术—计算机应用技术]