改进YOLOv5目标检测模型在城市街道场景中的应用

Application of improved YOLOv5 object detection model in urban street scenes

针对城市街道场景中小目标和密集目标检测准确率不高的问题,提出一种改进的YOLOv5的目标检测模型——City-YOLO。首先利用可变形卷积和Focus模块设计了两种新的下采样模块,使骨干网络(Backbone)与Neck层提取特征信息更准确,同时融合了全局特征信息;其次增加一层检测特征层,增强小目标的检测能力;随后将二阶通道注意力机制引入检测特征层,使网络更加关注目标的重要特征,提升网络对密集遮挡目标的检测能力;最后使用Kmeans算法重新设计了先验框,并将它分配到相应的检测特征层。在CityScape数据集上,City-YOLO模型的均值平均精度mAP50与mAP50:95分别达到了54.3%和30.6%;相较于YOLOv5,mAP50与mAP50:95分别提升了5.8和4.0个百分点。实验结果表明,City-YOLO模型有效地实现了高精度城市街道场景目标检测。

Aiming at the low accuracy problem of detecting small and dense objects in urban street scenes,an improved YOLOv5 object detection algorithm called City-YOLO was proposed.Firstly,two new down sampling modules were designed using deformable convolution and Focus module to enable the Backbone and Neck networks to extract information more accurately and integrate global feature information.Secondly,an additional detection feature layer was added to enhance the detection ability of small objects.Then,the second-order channel attention mechanism was introduced into the detection feature layer to make the network pay more attention to the important features of the object,improving the network detection ability for densely occluded objects.Finally,prior boxes were redesigned using K-means algorithm and assigned to the corresponding detection feature layers.For the CityScape dataset,the mAP50 and mAP50:95 of City-YOLO model are 54.3%and 30.6%respectively,compared with YOLOv5,which are improved by 5.8 and 4.0 percentage points respectively.The experimental results show that the City-YOLO model effectively achieves high-precision object detection in urban street scenes.