Dense Visual SLAM with Probabilistic Surfel Map.

Area of Science:

• Computer Vision
• Robotics
• Augmented Reality

Background:

• Visual SLAM is crucial for aligning virtual and real worlds in AR.
• RGBD dense Visual SLAM offers robustness and accuracy but faces challenges with measurement inconsistencies.
• These inconsistencies can degrade camera trajectory and scene reconstruction accuracy.

Purpose of the Study:

• To propose a novel map representation, the Probabilistic Surfel Map (PSM), for dense visual SLAM.
• To address RGBD measurement inconsistencies by encoding photometric and geometric uncertainties.
• To improve both camera pose estimation and scene reconstruction in dense visual SLAM.

Main Methods:

• Developed the Probabilistic Surfel Map (PSM) as a new global map representation.
• Modeled and updated sensor measurement uncertainties for improved accuracy.
• Applied uncertainty information to enhance front-end pose estimation and back-end optimization.

Main Results:

• Achieved significant reductions in Absolute Trajectory Error (ATE) and Relative Pose Error (RPE) compared to state-of-the-art methods like σ-DVO.
• Demonstrated a 40% ATE reduction and 18% RPE reduction in visual odometry.
• Showcased an 8.5% ATE reduction in complete SLAM and high-quality dense point cloud generation.

Conclusions:

• The Probabilistic Surfel Map (PSM) effectively mitigates RGBD inconsistencies in dense visual SLAM.
• PSM leads to substantial improvements in camera trajectory accuracy and scene reconstruction quality.
• This novel approach enhances the reliability and performance of AR applications.