复杂交通环境下目标检测中存在很多外界干扰因素,导致通用的目标检测算法效果较差.针对目标检测方法中全局特征信息利用不充分,小目标、遮挡目标检测精度低,以及模型计算量大等问题,提出一种基于改进YOLOv5s的融合全局特征目标检测方法.首先,对YOLOv5s的主干网络进行扩展,得到更深层的特征图以增强较大目标的语义信息;其次,在此基础上引入全局信息融合模块代替原模型中的Neck部分,以3D卷积的方式融合各尺度信息;然后,设计了一种基于位置的先验框匹配方法,在原图尺度上搜索与真实框匹配的先验框;最后,使用Copy-Paste数据增强方法增大小目标样本数量并使用DIoU-NMS作为后处理方法进行非极大值抑制.该模型在BDD100K数据集中平均精确率(mean Average Preci-sion,mAP)为54.55%,检测速度为63.72帧每秒(Frames Per Second,FPS).与原始YOLOv5s算法相比,该方法在检测速度及精度方面均有明显优势.
Object detection method in road scene based on fusion of global features
There are many external interference factors in object detection in complex traffic environment,which lead to poor detection effect of some general algorithms.A fusion global feature object detection method based on improved YO-LOv5s is proposed to address the insufficient utilization of global information,low accuracy of small object and occlusion object detection,and the large amount of calculation.Firstly,the backbone of YOLOv5s is expanded to obtain deeper fea-ture maps to enhance the semantic information of large objects.Secondly,a global information fusion module is intro-duced to replace the Neck part,and the information of each scale is fused by 3D convolution.Then,a location-based pri-or box matching method is designed,which searches for the prior box matching the bounding box on the original image scale.Finally,the Copy-Paste data augmentation method is applied to increase the number of small target samples,and DIoU-NMS is used as the post-processing method.The mean Average Precision(mAP)of the model in BDD100K dataset is 54.55%,the detection speed is 63.72 Frames Per Second(FPS).Compared with the YOLOv5s algorithm,this method has significant advantages in terms of detection speed and accuracy.
NETL
NSTL
万方数据
