引入Transformer的道路小目标检测
Road small object detection with Transformer
李丽芬 1黄如1
作者信息
- 1. 华北电力大学(保定)计算机系,河北保定 071003
- 折叠
摘要
针对道路场景中检测小目标时漏检率较高、检测精度低的问题,提出一种引入Transformer的道路小目标检测算法.在原YOLOv4算法基础上,对多尺度检测进行改进,把浅层特征信息充分利用起来;设计ICvT(improved convolu-tional vision transformer)模块捕获特征内部的相关性,获得上下文信息,提取更加全面丰富的特征;在网络特征融合部分嵌入改进后的空间金字塔池化模块,在保持较小计算量的同时增加特征图的感受野.实验结果表明,在KITTI数据集上,算法检测精度达到91.97％,与YOLOv4算法相比,mAP提高了 2.53％,降低了小目标的漏检率.
Abstract
Aiming at the problems of low detection accuracy and high missed detection rate when detecting small objects in road scenes,a road small object detection algorithm with Transformer was proposed.On the basis of the original YOLOv4 algorithm,the multi-scale detection was improved to make full use of the shallow feature information.The improved ICvT(improved convo-lutional vision Transformer)module was designed to capture the internal correlation of features,contextual information was obtained,and more comprehensive and rich features were extracted.The modified spatial pyramid pooling module was embedded in the feature fusion part of the network to increase the receptive field of the feature map while maintaining a small computational amount.Experimental results show that on the KITTI dataset,the detection accuracy of the algorithm reaches 91.97％,and compared with the YOLOv4 algorithm,the mAP is improved by 2.53％and the missing rate of small objects is reduced.
关键词
小目标检测/深度学习/YOLOv4算法/多尺度检测/Transformer/空间金字塔池化/特征融合Key words
small object detection/depth learning/YOLOv4 algorithm/multiscale detection/Transformer/spatial pyramid poo-ling/feature fusion引用本文复制引用
基金项目
国家自然科学基金项目(61906181)
中央高校基本科研业务费专项基金项目(2021MS094)
出版年
2024
NETL
NSTL
万方数据