High Accuracy Monocular SFM and Scale Correction for Autonomous Driving.

Area of Science:

• Robotics
• Computer Vision
• Artificial Intelligence

Background:

• Visual odometry is crucial for autonomous driving.
• Monocular systems face challenges with scale drift and accuracy compared to stereo systems.

Purpose of the Study:

• To develop a high-accuracy, real-time monocular visual odometry system for autonomous driving.
• To address scale drift and improve ground plane estimation.

Main Methods:

• Implemented robust Structure from Motion (SFM) using multithreading for challenging driving scenes.
• Utilized known camera height for scale correction.
• Developed a data-driven cue combination mechanism adapting observation covariances for accurate ground plane estimation.

Main Results:

• Achieved accuracy comparable to stereo systems and superior to prior monocular systems on the KITTI dataset.
• System outputs pose within 50 ms (worst-case) and operates over 30 fps (average).
• Significantly improved accuracy for ground plane-dependent applications like object localization.

Conclusions:

• The proposed monocular visual odometry system offers a viable, accurate, and efficient solution for autonomous driving.
• The novel cue combination method enhances ground plane estimation accuracy.
• The system demonstrates strong performance and potential for real-world applications.