DPO: Direct Planar Odometry with Stereo Camera.

Area of Science:

• Computer Vision
• Robotics
• Geometric Deep Learning

Background:

• Current direct visual odometry (VO) methods predominantly use points for camera pose estimation and environmental reconstruction.
• Direct Sparse Odometry (DSO) is a standard technique, but recent advancements include monocular plane-based approaches.
• Existing monocular plane-based DSOs have limitations, relying on potentially unreliable learned plane estimates or restricted to specific plane orientations.

Purpose of the Study:

• To propose the first Stereo Plane-based Visual Odometry (VO) system.
• To leverage planes as the primary features within a sliding window optimization framework, inspired by DSO.
• To evaluate the performance of the proposed plane-based stereo VO against established point-based methods.

Main Methods:

• Developed a novel Stereo Plane-based VO system utilizing planes exclusively as features.
• Integrated plane features into a sliding window optimization process.
• Employed dual quaternions for pose parameterization within the optimization.

Main Results:

• The proposed Stereo Plane-based VO demonstrates performance comparable to Stereo DSO, a leading point-based approach.
• The system effectively utilizes plane features for robust pose estimation in stereo visual odometry.
• Experimental results validate the efficacy of a purely plane-based feature approach in VO.

Conclusions:

• The novel Stereo Plane-based VO offers a viable alternative to point-based methods.
• This work advances plane-based approaches in visual odometry by extending them to stereo configurations.
• Future research can explore further integration of plane features for enhanced VO performance.