A novel approach to improve the precision of monocular visual odometry

Monocular visual odometry is an active research topic for mobile robot navigation due to its availability and simpleness. However, it inherently suffers from scale ambiguity inherently, so that the precesion of odometry becomes poor. In this paper, we propose a new method to resolve scale ambiguity for monocular visual odometry based on ground area extraction and a modified adaptive kalman-filter which is based on Support Vector Machine (SVM). Firstly, instead of using Homography directly, we combine the Watershed Algorithm with the proposed Edge Expansion Method (EEM) to realize the ground extraction. Secondly, for the purpose of reducing the possibility of divergency when using the kalman filtering algorithm in real scenes, this paper applies a modified SVM-based Adaptive Kalman Filtering algorithm (SVMAKF) to visual navigation area. We conduct some experiments in outdoor scenes to validate that these approaches can improve the accuracy of monocular visual odometry.