The Comparison of Different Visual Features for Visual Odometry

For robots to work effectively, the availability of a map with detailed information surrounding the workspace is an important requirement for indoor and outdoor tasks. This is usually achieved with using visual odometry techniques with feature-based methods. In this paper, we compare the performance of three different feature extraction methods: Scale Invariant Feature Transform (SIFT), Speed Up Robust Features (SURF) and Oriented FAST Rotated BRIEF (ORB). This paper presents experimental results on standard evaluation datasets and all experiments use measurement of the number of image correspondences as well as the ratio of good matched for the evaluation purpose. The results of experiments demonstrate that the performances of three methods in processing time, matching capability and accuracy. SIFT presents its stability in most scenarios although it is very slow. SURF is faster than SIFT and outperform SIFT on some scenarios. ORB is the most efficient feature and shows strong performance.