RC-SLAM: Road Constrained Stereo Visual SLAM System Based on Graph Optimization.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Intelligent vehicles often experience pose estimation errors in Visual Simultaneous Localization and Mapping (SLAM) systems.
• This is due to the discrepancy between the assumed six degrees of freedom (DoF) motion and the actual planar motion constrained by roads.

Purpose of the Study:

• To propose a stereo Visual SLAM system, RC-SLAM, that incorporates road constraints to enhance vehicle pose estimation accuracy.
• To address the challenge of representing road geometry and integrating it into the SLAM framework.

Main Methods:

• Approximating local roads as discrete planes and extracting parameters using homography.
• Establishing constraints between the vehicle and Local Road Planes (LRPs) to mitigate motion assumption errors.
• Employing epipolar constraints for robust rotation estimation, minimizing depth uncertainty impacts.
• Utilizing a nonlinear optimization model based on graph optimization for joint pose and map point optimization.

Main Results:

• The proposed RC-SLAM system demonstrated more accurate vehicle trajectory estimations compared to conventional methods.
• Experimental results on two datasets confirmed the effectiveness of incorporating vehicle-LRP constraints.
• Validation on a real-world dataset further substantiated the system's performance.

Conclusions:

• RC-SLAM effectively reduces pose estimation errors in Visual SLAM for intelligent vehicles by leveraging road geometry.
• The novel approach of using road plane constraints offers a robust solution for accurate vehicle localization in constrained environments.