A Probabilistic Feature Map-Based Localization System Using a Monocular Camera.

Area of Science:

• Robotics
• Computer Vision
• Probabilistic Modeling

Background:

• Image-based localization is a key technique in robotics and computer vision.
• Existing methods often require extensive image data and features for accurate localization.
• This research addresses the challenge of localization with insufficient image data.

Purpose of the Study:

• To develop a more accurate image-based localization method for scenarios with limited features.
• To improve camera pose estimation using a probabilistic map and 3D-to-2D matching.
• To enhance localization accuracy by optimizing initial pose estimates.

Main Methods:

• Generation of a probabilistic feature map incorporating sensor and camera pose uncertainties.
• Initial camera pose estimation using the conventional PnP algorithm on the probabilistic map.
• Optimization of the camera pose by minimizing Mahalanobis distance errors between map and query image features.

Main Results:

• The proposed method demonstrates improved localization accuracy compared to conventional algorithms.
• Validation was conducted through simulations and real-world experiments.
• The probabilistic map approach effectively handles environments with sparse features.

Conclusions:

• The developed probabilistic map-based localization method significantly enhances accuracy, especially in feature-scarce environments.
• The optimization technique refines initial pose estimates for robust localization.
• This approach offers a viable solution for challenging image-based localization tasks in robotics.