摘要
随着人工智能深度学习的快速发展,目标检测在智能视频监控、无人驾驶、交通管制等方面有着广泛的应用,尽管众多国内外的研究者在目标检测领域有一些突破,但是实际问题中目标的形变、遮挡以及光线变化等都是关键,那么如何设计合理的检测器适应不同的场景,提高模型的泛化能力也将是该领域的研究重点。论文具体场景是针对行人检测,因此在YOLOv3单阶段的目标检测基础上提出了一种用RepVGG替换主干网络的检测模型,该模型网络层数单一,并且采用了重参数化技术,而且在多尺度融合中将3个尺度的融合改成4尺度融合,提高模型的鲁棒性,在很好的拟合GPU的情况下,提高检测的精度和速度。
With the rapid development of artificial intelligence deep learning,target detection has a wide range of applications in intelligent video surveillance,unmanned driving,traffic control,etc.Although many domestic and foreign researchers have made some breakthroughs in the field of target detection,the target is actually a problem.The deformation,occlusion,and light changes are all key,so how to design a reasonable detector to adapt to different scenes and improve the generalization ability of the model will also be the focus of research in this field.The specific scenario in this article is for pedestrian detection,so based on YOLOv3single-stage target detection,a detection model that replaces the backbone network with RepVGG is proposed.The model has a single network layer and uses reparameterization technology.In the fusion,the fusion of 3 scales is changed to the fusion of 4 scales,which improves the robustness of the model,and improves the accuracy and speed of detection when the GPU is well fitted.
作者
何鹏元
马中
戴新发
夏静
HE Pengyuan;MA Zhong;DAI Xinfa;XIA Jing(th Research Institute,China State Shipbuilding Corporation Limited,Wuhan 430205)
出处
《舰船电子工程》
2022年第3期59-62,共4页
Ship Electronic Engineering