运动场景下小目标检测算法设计  

Design of Small Target Detection Algorithm in Motion Scenes

作　　者：李亚飞 刘娜[1] 周惠 杨雷 LI Yafei;LIU Na;ZHOU Hui;YANG Lei(School of Electrical Engineering and Automation,Tianjin University of Technology,Tianjin 300384,China;Autobrain(Tianjin)Technology,Co.,Ltd.,Tianjin 300300,China)

机构地区：[1]天津理工大学电气工程与自动化学院,天津300384 [2]奥特贝睿(天津)科技有限公司,天津300300

出　　处：《仪表技术》2024年第3期1-6,共6页Instrumentation Technology

基　　金：天津市自然科学基金青年项目(21JCQNJC00910);天津市自然科学基金重点项目(21JCZDJC00760);天津市“项目+团队”重点培养专项(XC202054)。

摘　　要：在运动场景下移动端设备常常面临图像模糊、质量低下以及目标体积小等问题,导致目标误检或漏检的情况频繁出现。同时,现有的模型参数量较大,无法满足实时性的要求。针对这些问题提出了YOLOv5s-C算法模型。首先,在主干网络(Backbone)引入坐标注意力机制,以增强模型对细节相关的通道特征的感知,从而提高模型定位和识别目标的能力,尤其对于模糊图像和小目标。其次,在特征加强网络(Neck)采用混合卷积GSConv和加权双向特征金字塔网络Bi-FPN,以获取全局上下文和不同尺度的信息,进而增强模型在图像模糊情况下对小目标的检测能力。最后,引入EIoU Loss作为边框回归损失函数,以加快模型的收敛速度和提升模型的检测精度。检测结果表明:在公开COCO2017数据集中,YOLOv5s-C算法模型参数量比原模型减少了29%,mAP@0.5∶0.95提升了1.8%,mAP@0.5提升了2.3%,小目标的误检、漏检情况得到了显著减少;在批量大小为32的情况下,该模型的速度达到了190.3 f/s。YOLOv5s-C算法模型在运动场景下的小目标检测中展现出优异的表现和广阔的应用前景。Mobile devices often face some problems such as image blur,low quality,and small target volume in sports scenes,leading to frequent false or missed detection of targets.Meanwhile,the existing models have a large number of parameters and cannot meet the requirements of real-time performance.A YOLOv5s-C algorithm model is proposed to address these issues.Firstly,a coordinate attention mechanism is introduced into the backbone network to enhance the models perception of detail related channel features,thereby improving the models ability to locate and recognize targets,especially for the blurred images and small targets.Secondly,in the feature enhancement network(Neck),a hybrid convolutional GSConv and a weighted bidirectional feature pyramid network Bi-FPN are used to obtain global context and information at different scales,thereby enhancing the models ability to detect small targets in image blur situations.Finally,EIoU Loss is introduced as the bounding box regression loss function to accelerate the convergence speed of the model and improve its detection accuracy.The test results show that in the publicly available COCO2017 dataset,the YOLOv5s-C algorithm model has a 29%reduction in parameter count compared to the original model,map@0.50.95 increased by 1.8%,map@0.5 improved by 2.3%,significantly reducing false positives and missed detections for small targets.In the case of a batch size of 32,the speed of the model reached 190.3 f/s.The YOLOv5s-C algorithm model has shown excellent performance and broad application prospects in small object detection in motion scenes.

关 键 词：模糊图像 小目标检测 坐标注意力机制 混合卷积 

分 类 号：TP391[自动化与计算机技术—计算机应用技术]