This project is carried out in the field of visual ego velocity estimation, which is simplification of visual motion estimation. The project aims to estimate the velocity using two types of cameras, one monocular and one stereo camera. Research investigated and found the steps needed to complete a feature-based motion estimation algorithm. The final monocular velocity estimation method uses a 2-D to 3-D motion estimation method where 3-D points are extracted using a Perspective-4-Points method. The physical distances between points are known previously and used to solve the well-known scale ambiguity problem of the monocular motion estimation. The implemented stereo method uses a 3-D to 3-D motion estimation method where 3-D points are found by matching KAZE features between left and right images, followed by triangulation. 3-D point matches over consecutive images are filtered asinliers or outliers based on mean absolute deviation.The methods were evaluated by installing a stereo camera onto an articulated arm robot, moving in a straight path in the forward direction of the camera. Ground truth values are extracted from the robot. Motion estimation values were extracted from the velocity estimation algorithms by inserting video files of the recorded robot motion. The monocular algorithm only considered footage from one of the stereo cameras two sensors.

Results show that the stereo method is less precise than the monocular case due to uncertainties in triangulated 3-D points. A better estimation was found when 3-D points are closer to the camera. The velocity estimated from the Monocular algorithm do have a consistent displacement between estimated and ground truth velocity. This was due to the physical points detected by the algorithm having inaccurate real-world measurements. The estimations of both methods improved when a higher velocity was applied. Then underlying reason was that the translation between images was increased, resulting in uncertainties being less significant.