Orthogonality Index Based Optimal Feature Selection for Visual Odometry

The performance of visual odometry is dependent upon the quality of features selected for computing the frame-to-frame transformation. In order to ensure the quality of selected features, conventional approaches consider the spatial distribution of the selected features, in addition to their counts and matching scores, in which a small number of features are selected randomly from each of the uniformly distributed buckets. In this paper, we show that features can be selected optimally, rather than randomly, using a well-defined mathematical formalism. The proposed method of optimal feature selection minimizes the degree of uncertainty in estimating the essential, fundamental, or homography matrix involved in visual odometry by maximizing the orthogonality index of individual equations and constraints associated with computation. We found that, at a constant noise level, the mean of the residual error and the variance of an estimated essential, fundamental, or homography matrix decrease monotonically with increasing orthogonality index. The simulation validates the increased accuracy of the feature selection based on the proposed orthogonality index compared with the conventional random selection. For instance, it enhances accuracy by as much as 35% when a small number of feature sets, say, 20 feature sets, are used. The experiments using the KITTI and Devon Island datasets further reinforce the performance enhancement of simulations by 9% and 20%, respectively.