Analysis and Reduction of Stereo Vision Alignment and Velocity Errors for Vision Navigation

Driverless vehicles have proved the effectiveness of VO (Visual Odometry) in a GPS (Global Positioning System)denied environment. VO is a DR (Dead Reckoning) system that provides position and alignment increments. A visual odometer can provide better performance than conventional odometers for wheeled robots and land vehicles on slippery grounds where the latter cannot estimate the speed correctly, but vision-based system has not been widely accepted as a reliable navigation system due to several issues. First it suffers from various error sources in images such as noise, motion blur and distortion. The visual measurements are also ambiguous in featureless, self-similar or dynamic environments, or during rapid motion which causes many mismatches in the corresponding image points. Last but not least, the increment error in position and alignment will accumulate over time. In this paper, we are going to investigate how long stand-alone stereo cameras can provide acceptable navigation solution for car navigation during GPS outages. This paper will investigate the use of stereo cameras to overcome scaling ambiguity in image observations, assess the accuracy of stereo vision alignment and velocity determination and develop methods to reduce alignment and position increments errors. Windowed camera pose estimation, Bucketing, (BA) Bundle adjustment and loop closure were used to reduce position and alignment error. We compared 3D-3D and 3D-2D motion estimation method and we also compared tracking and matching method. Our results have shown considerable position and alignment error reduction using windowed camera pose estimation. Bucketing guarantees that all the feature points are well distributed along z axis and it reduces the computational complexity. BA improves the results as well. Loop closure helps to considerably reduce the position and alignment errors in the case of revisited areas. 3D-3D VO drifts considerably faster than 3D-2D method and tracking gives faster and more accurate results than matching method.