Photogrammetric Feature Optimization and Blunder Detection for Stereo Visual Odometry

초록

Stereo visual odometry is the technique to estimate platform motion using stereo images. This technique extracts features within an image sequence and uses them to estimate the motion. Therefore, feature extraction affects the overall accuracy. It is important for visual odometry to select and use optimal features. In this study, we applied a new feature optimization and blunder detection method to visual odometry. We aimed to apply photogrammetric and statistical processing to improve the accuracy of visual odometry. The developed visual odometry is performed in four steps: feature extraction and tracking, feature optimization, motion estimation, blunder detection and motion reestimation. In the feature extraction and tracking step, the Shi-Tomasi corner method and the KLT (Kanade-Lucas- Tomasi) method are carried out for real-time image processing. In the feature optimization step, outliers are detected and eliminated. This step is divided into photogrammetry-based and vision-based processing. In the photogrammetrybased processing, the projection and re-projection errors of features are calculated using the coplanarity condition. By setting a tolerance for the errors, features are classified as inliers or outliers. In the vision-based processing, the RANSAC (random sample consensus)-based filtering is performed within multiple images. In the next step, the optimized features are used for bundle adjustment based the collinearity condition and platform motion is estimated. After motion estimation, the features are examined by blunder detection. In this step, the errors of all features are checked and the features with gross errors are reclassified as outliers. The motion is re-estimated using the final inliers. For experiments, we used the dataset provided by KITTI (Karlsruhe institute of technology and Toyota technological institute) community. We analysed the result of the feature optimization and blunder detection and the effects of the photogra