Visual Odometry Using Feature Point and Ground Plane for Urban Environments

In this paper, we propose a visual odometry system for accurate and robust 6DOF localization in urban environments. The visual odometry system integrates a feature based Visual odometry and normal vector information of a estimated ground plane in order to achieve the accurate Visual Odometry in untextured environments like urban. In the feature-based Visual Odometry, robustness and accuracy are improved by using RANSAC, three-point algorithm, and key frame adjustment. In the ground plane extraction, the stereo homography matrix is used. We use RANSAC to calculate the homography matrix robustly, . Unscented Kalman Filter is used for the sensor fusion to integrate these information. Normalized innovation squared Test is used for the fault detection of the ground plane extraction. In addition, we propose a localization system which integrates the Visual odometry system and a gyro sensor to reduce the influence of dynamic objects. Localization results of our Visual odometry system and localization system are presented, the effectiveness of our system is demonstrated.