Scholarlink

<p>This study evaluates ten visual odometry (VO) methods, comparing five traditional detector-based techniques, such as SIFT and ORB, with five deep learning-based approaches, including SuperPoint, DISK, ALIKED, D2Net, and XFeat. Two matching strategies, Brute-Force and LightGlue, are applied to assess their performance on the KITTI dataset, using both raw images and perturbed inputs with resolution downscaling, blurring, and noise. The evaluation is based on seven key metrics: translation error, rotation error, absolute trajectory error, relative pose error (distance, rotation), cumulative drift, and frames per second. Experimental results show that traditional methods provide better stability and real-time performance, while deep learning-based methods demonstrate superior accuracy in certain segments but tend to be more sensitive to strong image perturbations. This study highlights the trade-off between robustness and representational power, suggesting that, although deep learning-based approaches still require improvements in generalization, they hold strong potential for future VO systems. Additionally, a standardized evaluation framework is proposed to assist in selecting appropriate algorithms for real-world applications.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Deep Visual Odometry: A Comprehensive Survey on Performance Analysis of Feature Detection and Matching Techniques

  • 期刊论文

  • 发表日期：

    2026年2月20日

查看全文
  • Ji Hun Choi,
  • Chang Ho Kang,
  • Sun Young Kim

摘要

This study evaluates ten visual odometry (VO) methods, comparing five traditional detector-based techniques, such as SIFT and ORB, with five deep learning-based approaches, including SuperPoint, DISK, ALIKED, D2Net, and XFeat. Two matching strategies, Brute-Force and LightGlue, are applied to assess their performance on the KITTI dataset, using both raw images and perturbed inputs with resolution downscaling, blurring, and noise. The evaluation is based on seven key metrics: translation error, rotation error, absolute trajectory error, relative pose error (distance, rotation), cumulative drift, and frames per second. Experimental results show that traditional methods provide better stability and real-time performance, while deep learning-based methods demonstrate superior accuracy in certain segments but tend to be more sensitive to strong image perturbations. This study highlights the trade-off between robustness and representational power, suggesting that, although deep learning-based approaches still require improvements in generalization, they hold strong potential for future VO systems. Additionally, a standardized evaluation framework is proposed to assist in selecting appropriate algorithms for real-world applications.