A Generic Image Processing Pipeline for Enhancing Accuracy and Robustness of Visual Odometry.

Area of Science:

• Robotics
• Computer Vision
• Image Processing

Background:

• Visual Odometry (VO) accuracy is challenged by environmental factors like lighting variations and feature matching outliers.
• Feature-based VO algorithms are susceptible to performance degradation due to these factors.

Purpose of the Study:

• To propose a generic and modular image processing pipeline to enhance the accuracy and robustness of feature-based VO algorithms.
• To address lighting conditions, feature distribution, and outlier rejection in VO.

Main Methods:

• Implemented Contrast Limited Adaptive Histogram Equalization (CLAHE) for lighting normalization.
• Utilized the Suppression via Square Covering (SSC) algorithm for improved feature distribution.
• Introduced the Angle-based Outlier Rejection (AOR) algorithm for robust outlier removal.

Main Results:

• The pipeline demonstrated significant improvements in VO accuracy and robustness across various datasets (KITTI, TUM) and robot platforms.
• Validated effectiveness for monocular, RGB-D, and stereo VO configurations.
• Achieved performance gains without compromising computational efficiency.

Conclusions:

• The proposed pipeline effectively enhances feature-based VO performance by mitigating common issues.
• It offers a versatile solution applicable to diverse VO systems and environments.
• The method provides a substantial accuracy and robustness improvement over standard VO approaches.