Indoor Visual Positioning Aided by CNN-Based Image Retrieval: Training-Free, 3D Modeling-Free

Area of Science:

• Computer Vision
• Robotics
• Geographic Information Systems

Background:

• Indoor localization is crucial for location-based services (LBS) like navigation and robotics.
• Existing visual localization systems face challenges balancing accuracy and cost.
• Robust and efficient indoor visual localization remains an open problem.

Purpose of the Study:

• To propose a novel indoor localization method addressing the accuracy-cost trade-off.
• To leverage Convolutional Neural Network (CNN) features for robust image retrieval and pose estimation.
• To enable accurate geo-localization using only RGB images and reference poses.

Main Methods:

• A two-part approach: CNN-based image retrieval and pose estimation.
• Utilizing pre-trained deep convolutional neural networks (DCNNs) for feature extraction and similarity comparison.
• Employing a monocular visual odometer-inspired scheme for pose estimation with lightweight scene representation.

Main Results:

• The proposed scheme achieves high location accuracy and orientation estimation.
• Demonstrated efficiency and applicability in complex indoor environments.
• Enhanced positioning accuracy and usability compared to similar solutions.

Conclusions:

• The developed method offers an efficient and accurate solution for indoor visual localization.
• The approach is adaptable to both indoor and outdoor environments.
• Compatible algorithms for data acquisition and pose estimation support future data expansion.