福建电脑2025，Vol.41Issue(9)：16-23,8.DOI:10.16707/j.cnki.fjpc.2025.09.004
DFA-CenterNet无锚框人车检测方法
DFA-CenterNet Human-Vehicle Detection Method without Anchor Frame
倪健 1王毅飞 1关帅鹏2
作者信息
- 1. 河北工程大学信息与电气工程学院 河北 邯郸 056038
- 2. 中国船舶集团有限公司第七一八研究所 河北 邯郸 056000
- 折叠
摘要
Abstract
This paper proposes an improved detection algorithm,DFA CenterNet,based on an anchor free frame architecture to address the issues of high missed detection rates and insufficient multi-scale generalization ability in complex traffic scenarios.Adopting a dense bidirectional feature fusion module,multi-resolution feature complementarity is achieved through feature superposition and differential dual branch structure.Combined with convolutional attention and spatial enhancement mechanism,the feature weight distribution is dynamically optimized in the channel and spatial dimensions to suppress background noise interference in occluded areas.The mAP of this method on the test set is 89.76%,which is 10.31%higher than the original CenterNet,and the real-time detection speed is 40.79 FPS.The experimental results show that the proposed method can significantly improve the detection robustness of multi-scale human vehicle targets,providing a feasible solution for intelligent perception tasks in complex traffic scenarios.关键词
目标检测/无锚框方法/注意力机制/交通场景Key words
Object Detection/No Anchor Box Method/Attention Mechanism/Traffic Scene分类
信息技术与安全科学引用本文复制引用
倪健,王毅飞,关帅鹏..DFA-CenterNet无锚框人车检测方法[J].福建电脑,2025,41(9):16-23,8.
