Incorporating a wheeled vehicle model in a new monocular visual odometry algorithm for dynamic outdoor environments

Area of Science:

• Robotics
• Computer Vision
• Autonomous Systems

Background:

• Monocular visual odometry (VO) is crucial for autonomous navigation.
• Existing VO methods often struggle with dynamic environments and complex vehicle motion.
• Accurate estimation of vehicle motion parameters like yaw rate and side slip angle is essential.

Purpose of the Study:

• To develop an advanced monocular VO algorithm for ground vehicles.
• To improve the robustness and accuracy of visual odometry by incorporating a wheeled vehicle model.
• To address limitations of the planar-motion hypothesis in real-world scenarios.

Main Methods:

• Utilizes a single-track bicycle model to relate yaw rate and side slip angle.
• Incorporates pitch angle estimation to account for real-world dynamics.
• Employs linearization and a RAndom SAmple Consensus (RANSAC) scheme for efficient motion parameter estimation.
• Refines solutions by minimizing reprojection error using inlier data.

Main Results:

• The proposed algorithm demonstrates improved accuracy in estimating vehicle motion parameters.
• It achieves robust performance in dynamic outdoor environments over several kilometers.
• Outperforms existing state-of-the-art monocular VO methods on synthetic and real-world datasets.

Conclusions:

• The integrated wheeled vehicle model significantly enhances monocular VO performance.
• The algorithm is suitable for real-time on-board visual localization applications.
• This approach offers a more accurate and reliable solution for autonomous vehicle navigation.