Real-Time Photometric Calibrated Monocular Direct Visual SLAM

Area of Science:

• Computer Vision
• Robotics
• Photogrammetry

Background:

• Mobile ground equipment often faces challenges with illumination sensitivity, impacting the accuracy of simultaneous localization and mapping (SLAM).
• Existing visual SLAM algorithms struggle to maintain robustness under varying light conditions.

Purpose of the Study:

• To develop an enhanced visual SLAM algorithm addressing illumination sensitivity issues for mobile ground equipment.
• To improve the accuracy and robustness of SLAM systems in dynamic lighting environments.

Main Methods:

• Optimized camera photometric formation, including vignette and response functions.
• Employed Shi-Tomasi corner tracking and sparse direct visual odometry (VO) for pixel tracking and optimization.
• Utilized the Levenberg-Marquardt (L-M) method for joint optimization and real-time photometric calibration parameter updates.
• Implemented a Shi-Tomasi corner filtering strategy to reduce computational complexity and ORB features for loop closure detection.

Main Results:

• The proposed algorithm demonstrated real-time dynamic compensation of exposure, effectively reducing the impact of light variations.
• Experimental results on TUM, KITTI, EuRoC datasets, and an actual environment indicate promising positioning and mapping performance.
• The enhanced SLAM algorithm shows improved accuracy and robustness compared to standard methods under varying illumination.

Conclusions:

• The developed enhanced visual SLAM algorithm effectively mitigates illumination sensitivity problems in mobile ground equipment.
• The dynamic photometric calibration and optimized feature tracking contribute to robust and accurate localization and mapping.
• The algorithm presents a viable solution for real-world applications requiring reliable navigation under changing light conditions.