A Classification-Based Visual Odometry Approach

Monocular visual odometry is a well-studied, but still unresolved problem. In recent years. Most of the successful monocular visual odometry systems are feature based approaches. However, not all feature points are beneficial for the matching step. As these feature points are extracted for repeatability and rotation invariance, they are not necessarily suitable for the matching process. In this paper, we propose a novel approach for monocular visual odometry, in which we choose the ORB as our feature to meet the real-time requirement, and introduce a random fern classification step before the matching step. The random fern classifier could help us discard those unmatchable feature points, thus speeding up the matching step and enable us to obtain a better visual odometry performance. Experiments on a dataset containing around 2400 images confirm the validity of our proposed approach.