Related Experiment Video
Updated: Aug 13, 2026
08:25
Combining Eye-tracking Data with an Analysis of Video Content from Free-viewing a Video of a Walk in an Urban Park Environment
Published on: May 7, 2019
An efficient algorithm for learning-based visual localization
Jindi Zhong1, Ziyuan Guo1, Hongxia Wang1
1College of Electrical Engineering and Automation, Shandong University of Science and Technology, No. 579, Qianwangang Road, Qingdao, 266590, China.
Summary
This study introduces a new optimization algorithm for visual localization in GPS-denied areas. The method enhances deep neural network training and speeds up convergence, improving accuracy.
Area of Science:
- Robotics and Computer Vision
- Optimization Algorithms
- Deep Learning
Background:
- Visual localization is crucial for navigation in environments lacking Global Positioning System (GPS) signals.
- Existing methods struggle with precision and robustness in GPS-denied scenarios.
- Deep neural networks (DNNs) are increasingly used for visual localization but require efficient training.
Purpose of the Study:
- To develop a novel optimization algorithm for high-precision and robust visual localization.
- To improve the training efficiency and convergence speed of DNNs for localization tasks.
- To address the challenges of localization in GPS-denied environments.
Main Methods:
- Proposed a new optimization algorithm based on the optimal control principle (OCP).
- Incorporated diagonal Hessian information estimation to exploit curvature information.
- Theoretically analyzed the algorithm's convergence rate, achieving O(1T).
Main Results:
- Demonstrated significant improvements in visual localization accuracy on public datasets.
- Achieved a 33.71% increase in position accuracy and 15.66% in rotation accuracy compared to Adam on the Great Court scene.
- Showcased strong generalization capabilities across various tasks.
Conclusions:
- The proposed OCP-based optimization algorithm offers a superior approach for visual localization in GPS-denied environments.
- The method effectively enhances DNN training efficiency and accelerates convergence.
- Experimental results validate the algorithm's effectiveness and generalization ability.
