| 摘 要: 传统闭环检测算法采用人工设计特征,准确性差且特征提取耗时,难以满足实时性需求。因此,提出了一种基于改进Super Point与多尺度描述符融合的视觉同时定位与建图 SLAM(Simultaneous Localization And Mapping)闭环检测算法,旨在提高系统的准确性并满足实时性需求。该算法引入残差结构替代原卷积结构,同时融合了多尺度特征融合策略,最后通过Light Glue网络进行特征匹配,利用New College数据集对算法进行验证。在准确率为100%的情况下,改进算法的最大召回率在左右两个数据集上分别能达到36.4%和34.8%,满足闭环检测需求。 |
| 关键词: 闭环检测 特征提取 多尺度特征融合 特征匹配 |
|
中图分类号:
文献标识码: A
|
| 基金项目: 国家自然科学基金资助项目(61663005) |
|
| Visual SLAM Loop Closure Detection Algorithm Based on Improved Super Point and Multi-Scale Fusion |
|
ZHANG Lili1, XU Ming1,2, MA Longhua2, TIAN Guanzhong3
|
(1.School of Information Science and Engineering, Zhejiang Sci-Tech University, Hangzhou 310018, China;
2.School of Information Science and Engineering, Ningbo Tech University, Ningbo 315199, China;
3.Ningbo Global Innovation Center Zhejiang University, Ningbo 315199, China)
lilizhang032@163.com; photoman2004@126.com; lhma_zju@zju.edu.cn; gztian@zju.edu.cn
|
| Abstract: Traditional loop closure detection algorithms rely on handcrafted features, which are less accurate and time-consuming in feature extraction, making it difficult to meet rea-l time requirements. This paper proposes a visual
(Simultaneous Localization and Mapping,SLAM)loop closure detection algorithm based on improved SuperPoint and mult-i scale descriptor fusion, aiming to enhance the system’s accuracy and meet rea-l time demands. The algorithm
replaces the original convolutional structure with a residual structure and incorporates a mult-i scale feature fusion strategy. Finally, feature matching is performed using the LightGlue network. The algorithm was validated on the
NewCollege dataset. At 100% precision, the improved algorithm achieves maximum recall rates of 36.4% and 34.8% on the left and right datasets, respectively, meeting the requirements for loop closure detection. |
| Keywords: loop closure detection feature extraction mult-i scale feature fusion feature matching |
