Integrating visual odometry and dead-reckoning for robot localization and obstacle detection

The research presented introduces a new methodology to infer environment structure and robot localization by using a monocular machine vision system. The local field of view is constrained to the vicinity of the mobile robot in order to accomplish with robust navigation issues. The strategy proposed uses optical flow techniques and planar models to obtain qualitative 3D information and robot localization by using time integration series of acquired frames. In this way, different space resolution and meaningful corner information are considered. The different significant image points are correlated from camera pose knowledge and odometer data. The robot localization is achieved combining on board and visual odometer systems for reducing dead reckoning problems. Therefore, the two system errors are compared in a parallel process that selects the most accurate robot localization. Moreover, it is used to infer qualitative 3D information, when mismatches between both odometer systems are produced, due to the fact that the planar floor model is not accomplished. In this context, experimental results that reinforce the effectivity of the work developed are reported by using the available lab mobile platform. Other remarkable features of the strategy presented are its simplicity and the low computational cost.