WPO-Net: Windowed Pose Optimization Network for Monocular Visual Odometry Estimation.

Area of Science:

• Computer Vision
• Robotics
• Artificial Intelligence

Background:

• Visual odometry is crucial for autonomous driving, estimating camera pose in 3D space.
• Existing learning-based methods offer robustness but may still require camera calibration.
• The need for calibration-free, noise-robust visual odometry methods persists.

Purpose of the Study:

• To propose a novel, calibration-free method for estimating the 6 Degrees of Freedom (6DoF) pose of a monocular camera.
• To introduce the "windowed pose optimization network" for enhanced visual odometry.
• To evaluate the proposed method's performance against state-of-the-art techniques.

Main Methods:

• Developed a supervised learning-based network architecture featuring a feature encoder and pose regressor.
• Utilized multiple consecutive grayscale image pairs for training.
• Enforced composite pose constraints during the learning process.

Main Results:

• Achieved a rotational error of 3.12 degrees per 100 meters on the KITTI dataset.
• Demonstrated efficient training time (41.32 ms) and inference time (7.87 ms).
• Exhibited competitive performance compared to existing state-of-the-art methods.

Conclusions:

• The proposed "windowed pose optimization network" offers a novel and effective calibration-free approach to visual odometry.
• The method shows significant potential for improving the robustness and accuracy of localization in autonomous driving systems.
• The achieved performance metrics highlight the technique's viability and novelty.