Visual Odometry Based on Gabor Filters and Sparse Bundle Adjustment

The optimal way for recovering motion and structure from long image sequences is to use sparse bundle adjustment. The objective of this work was to elaborate a method yielding good initial estimates of the pose for SBA based pose refinement. A new approach for determining inter frame correspondences between features is presented. It is based on singular value decomposition of weighted cross correlation matrix of two feature sets. The weighting of potential matches between features is realized on the basis of cross correlation of intensity values and Gabor filter responses. The method is robust to large motions. Corner and SIFT features were used to compare the effectiveness of our method with Kalman filter based tracking of features. Initial estimate of the pose is determined with singular value decomposition and quaternion based representation of camera´s pose. The refinement of the pose estimate is achieved using RANSAC and then SBA on several consecutive frames at once. Experimental results demonstrate the capability of the system to estimate visual odometry in real-time.