曲宸阳,程艳云.基于改进YOLOv7的交通标志检测算法[J]. 微电子学与计算机,2024,41(7):8-17. doi: 10.19304/J.ISSN1000-7180.2023.0399
引用本文: 曲宸阳,程艳云.基于改进YOLOv7的交通标志检测算法[J]. 微电子学与计算机,2024,41(7):8-17. doi: 10.19304/J.ISSN1000-7180.2023.0399
QU C Y,CHENG Y Y. Traffic sign detection algorithm based on improved YOLOv7[J]. Microelectronics & Computer,2024,41(7):8-17. doi: 10.19304/J.ISSN1000-7180.2023.0399
Citation: QU C Y,CHENG Y Y. Traffic sign detection algorithm based on improved YOLOv7[J]. Microelectronics & Computer,2024,41(7):8-17. doi: 10.19304/J.ISSN1000-7180.2023.0399

基于改进YOLOv7的交通标志检测算法

Traffic sign detection algorithm based on improved YOLOv7

  • 摘要: 随着智能驾驶系统飞速发展,交通标志检测技术受到广泛关注。针对交通标志在图像中像素面积小、分辨率低、背景复杂等问题,提出了一种基于改进YOLOv7的交通标志检测算法。首先,构建增强特征提取模块。采用残差瓶颈结构和全维度动态卷积层优化特征提取网络中可拓展高效层聚合网络结构,不仅提高了特征提取网络聚焦小目标交通标志关键特征的能力,而且还避免了特征丢失。其次,在特征融合网络中嵌入轻量型混合注意力模块,过滤小目标交通标志周围复杂背景噪声,使网络的颈部更好地融合浅层细节信息和深层语义信息,增强多尺度特征融合效果。最后,解耦网络检测头使用两条享有不同参数的独立分支分别完成小目标交通标志分类和回归任务,提高分类回归准确度。在TT100K交通标志检测数据集上进行了实验评估,结果表明:相较于基线YOLOv7算法,改进算法的小目标精度提高了1.9%、小目标召回率提高了3.1%、mAP值提高了2.6%;同时,改进算法检测速度为57.1 帧/s,满足实时检测的要求。

     

    Abstract: With the rapid development of intelligent driving system, traffic sign detection technology has been widely concerned. Aiming at the problems of small pixel area, low resolution and complex background of traffic signs in images, an improved traffic sign detection algorithm based on YOLOv7 is proposed. Firstly, an enhanced feature extraction module is constructed, which uses the residual bottleneck structure and the full-dimensional dynamic convolution layer to optimize the extendable efficient layer aggregation network structure in the feature extraction network, so as to improve the ability of the feature extraction network to focus on the key features of small-target traffic signs and avoid feature loss. Secondly, a lightweight hybrid attention module is embedded in the feature fusion network to filter the complex background noise around small target traffic signs, so that the neck of the network can better integrate the shallow detail information and deep semantic information, and enhance the effect of multi-scale feature fusion. Finally, the network detection head is decouped, and two independent branches with different parameters are used to complete the classification and regression tasks of small target traffic signs respectively, which improves the accuracy of classification and regression. Experimental evaluation is conducted using the TT100K traffic sign detection data set. The results show that compared with baseline YOLOv7 algorithm, the accuracy of small target is increased by 1.9%, the recall rate of small target is increased by 3.1%, and the mAP value is increased by 2.6%. Additionally, the improved algorithm achieves a detection speed of 57.1 frames per second, meeting the requirements of real-time detection.

     

/

返回文章
返回