Stereo Vision Based Ego-Motion Estimation with Sensor Supported Subset Validation

We propose a method to reliably estimate the motion of a dynamic stereo camera system in the three dimensional world where observations are disturbed by high portions of independently moving objects. Robustness of the estimation process is achieved by applying an additional visual sensor. The system consists of a stereo vision sensor, acquiring optical flow and depth information of the scene and a camera with its optical axis oriented perpendicular to the road surface, measuring the speed over ground of the camera-equipped vehicle. The fusion approach presented in this paper combines the motion estimates of the two sensors and applies an efficient random sampling scheme that evaluates the distribution of motion patterns in the scene. The goal of the sampling scheme is to separate the observations into alien and ego-motion portions used in the subsequent step to extract the ego-motion of the camera system. The presented setup of the two visual sensors in combination with the observation sampling scheme increases robustness of the overall system.