Egomotion estimation for monocular camera visual odometer

Abstract

Visual odometer is a technique that permit to measure the velocity of a vehicle by the elaboration of a video source from a camera solidly fixed to it. Actual cameras are cheap and have high resolution and fast frame rate, with these characteristics it become really easy to estimate the ego motion with a really high accuracy. It is possible to achieve a resolution in position around 20–30 cm. The use of VO as part of the navigation system in autonomous vehicles is nowadays a field with a big interest for the science and industry community. The aim is not to substitute the current navigational systems such as GPS or INS, but to provide robustness and safety to operations, for example when there is a loss of signal coverage. Extended work has been done using stereo camera, but recently the use of a monocular camera has become more popular. A research about the concept and related work is covered in this work. Also the implementation of a simple visual odometer (VO) to recover the speed of an on-road car with straight motion with a monocular camera is presented in this document. This algorithm is based on the detection of a region of interest within the road and the computation of the relative motion of this region. Three different methods have been tested to estimate this relative motion: optical flow, feature matching and feature tracking. Finally the best method is used to test the algorithm with a real traffic sequence. The results are compared and modifications are proposed to improve it.