改进YOLOv4-tiny模型的交通图像目标检测  

Traffic Image Target Detection Based on Improved Yolov4 Tiny Mode

作　　者：王竣生 WANG Junsheng(School of Computer Science,Liupanshui Normal University,Guizhou,China,553000)

机构地区：[1]六盘水师范学院计算机科学学院,贵州553000

出　　处：《福建电脑》2022年第11期13-18,共6页Journal of Fujian Computer

基　　金：六盘水师范学院校级基金项目(No.LPSSYZK202011)资助。

摘　　要：道路交通中的实时智能安全监控系统拥有着海量图像数据源。数据源中具有不同路况、不同天气、多类别车型和多种分辨率的道路交通图像数据。针对该类图像,本文提出了一种基于YOLOv4-tiny的轻量级目标检测模型,以快速捕捉目标车辆的行驶信息。首先,采用移动翻转瓶颈卷积对骨干结构进行优化,有助于实现更好的内存效率;其次,在模型的颈部网络中使用改进的空间金字塔池化结构,将多尺度的局部特征连接在同一卷积层中,从而增加局部区域特征图的接受域;最后,在网络中添加一个尺度层,将顶层的特征图合并,获得细粒度的特征,提高了检测精度,特别是对小目标的检测。实验结果表明,与以往模型相比,该模型结构具有较高的精确度,需要的存储空间最小,能够高效检测并提取空间内的目标信息,实现智能监控。The real-time intelligent safety monitoring system in road traffic has a large number of image data sources, which contain road traffic image data of different road conditions, different weather, multiple types of vehicles and multiple resolutions. For this kind of image, a lightweight target detection model based on yolov4tiny is proposed to quickly capture the driving information of the target vehicle. Firstly, mobile inverted bottleneck convolution is used to optimize the backbone structure, which helps to achieve better memory efficiency;Secondly, the improved spatial pyramid pooling structure is used in the neck network of the model to connect the multi-scale local features in the same convolution layer, thereby increasing the acceptance domain of the local regional feature map. Third, add a scale layer in the network, merge the top-level feature maps to obtain fine-grained features, which improves the detection accuracy, especially for small targets. The experimental results show that compared with the previous models, the model structure has higher accuracy, requires the smallest storage space, and can effectively detect and extract the target information in the space to achieve intelligent monitoring.

关 键 词：YOLOv4-tiny 移动翻转瓶颈卷积 空间金字塔池化 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]