«上一篇/Previous Article|本期目录/Table of Contents|下一篇/Next Article»

《计算机技术与发展》[ISSN:1006-6977/CN:61-1281/TN]

卷:

34

期数:

2024年07期

页码:

131-137

栏目:

人工智能

出版日期:

2024-07-10

文章信息/Info

Title:

Feature Point Detection and Visual Location Based on Distinctiveness Evaluation

文章编号:

1673-629X(2024)07-0131-07

作者:

王欢1; 李创2

1. 广东省科学技术情报研究所,广东 广州 510033; 2. 西安交通大学 人工智能学院,陕西 西安 710048

Author(s):

WANG Huan1; LI Chuang2

1. Guangdong Institute of Scientific & Technical Information,Guangzhou 510033,China; 2. School of Artificial Intelligent,Xi'an Jiaotong University,Xi'an 710048,China

关键词:

人工智能; 特征点检测; 深度学习; 视觉里程计; 同时定位与构图

Keywords:

artificial intelligence; feature point detection; deep learning; visual odometry; simultaneous localization and mapping

分类号:

TP183

DOI:

10.20165/j.cnki.ISSN1673-629X.2024.0128

摘要:

传统的特征点检测算法难以应对实际场景中的光照和视点变化,基于深度学习的特征点检测算法得到的特征点 的定位精度不足,且难以剔除局部区域的相似特征点。 针对基于深度学习特征点提取面临的问题,设计了基于特征融合和独特性评价的特征点检测算法。 首先,为提高特征点的定位精度,采用基于特征融合的网络结构以及对应的特征融合损失函数,解决高层特征中细节特征偏移以及模糊的问题。 其次,将特征点是否来自局部相似区域转换为对特征点的独特性评价,在网络结构中增加独特性分支并设计独特性损失函数以学习特征点的独特性响应值。 通过提取独特性响应值较高的特征点,剔除局部相似区域的特征点以减少后续特征匹配中误匹配的数量。 采用视觉里程计和视觉同时定位与构图系统对算法进行了验证,在 KITTI 和 TUM 数据集上验证了算法在大范围室外场景和小范围室内场景下均具有良好的 鲁棒性和定位性能。

Abstract:

Traditional feature point detection algorithms are difficult to cope with the changes in lighting and viewpoint in the real scenes.The positioning accuracy of feature points obtained by the feature point detection algorithm based on deep learning is insufficient,and it is difficult to eliminate similar feature points in local areas. In order to improve the performance of algorithm based on deep learning,a feature point detection algorithm based on feature fusion and uniqueness evaluation is designed. Firstly, in order to improve the positioning accuracy of feature points,the network structure based on feature fusion and the corresponding feature fusion loss function are used to solve the problems of detail feature offset and blur in high-level features. Secondly,whether the feature points are from locally similar regions is converted into the uniqueness evaluation of the feature points,the uniqueness branch is added to the network structure,and the uniqueness loss function is designed to learn the uniqueness response value of the pixels in the predicted image. By extracting feature points with high uniqueness response values,the feature points of locally similar regions are excluded to reduce the number of mis-matched in subsequent feature matches. Based on the algorithm,the visual odometry and visual simultaneous localization and mapping system are constructed,and the system has good robustness and accurate positioning ability in large-scale outdoor scenes and small-scale indoor scenes on the KITTI and TUM datasets.

常用功能

更新日期/Last Update: 2024-07-10