3D map building for a humanoid robot by using visual odometry

Localization and 3D mapping are important tasks, for a humanoid robot to move in a complex human environment. In This work, we present a feature tracking based visual odometry system using stereo vision to achieve simultaneous 6D localization and 3D mapping. A camera position can be estimated from 3D environmental depth data and a previous 3D map, and used to incrementally update the 3D map. Visual odometry was employed to compute the view transformation relating a pair of views. Experimental results that estimate camera trajectory and build a 3D map are accurately obtained. The system is more accurate than typical dead-reckoning system such as gyro sensors.