Visual odometry for road vehicles using a monocular camera
Henrik Berg, Raman Haddad · Chalmers Publication Library (Chalmers University of Technology) · 2016
As technology keeps advancing, autonomous vehicles are more than a possibility, they are inevitable.An inevitability offering both security and comfort for passengers as well as for others in traffic.However there are still certain predicaments to be investigated, one of those challenges involves an accurate positioning system of the vehicle, especially when GPS is not available.One method dealing with this issue is to incrementally estimate motion using images taken by a digital camera, an area know as Visual Odometry.However Visual Odometry algorithms can be implemented in many different ways, hence there is a need to evaluate the performance on real data.The aim of the thesis is to compare the different single camera Visual Odometry algorithms with respect to vehicle trajectory (the rotational and translational error) and the execution time.The algorithms differed with respect to the used feature detectors and descriptors and the feature matching/tracking method.The investigated feature detectors, descriptors and tracking were based on FAST/ORB, SIFT and SURF methods and compared with Kanade-Lucas-Tomasi (KLT) tracking.The relative motion between two consecutive images was estimated from 2D-to-2D feature correspondence and Nister's five-point algorithm.The algorithms are implemented in C++/OpenCV and tested on three image sequences in different environments from the public KITTI dataset.The obtained results are compared to ground truth data from a highly accurate GPS.The results show that the investigated methods are able to estimate the ego-motion with an average translation error of <7 % and a rotation error of <0.02 deg/m.The best results, with respect to rotational and translational error, are obtained using feature matching of SIFT features along with the corresponding descriptor.The results also show that the feature tracking using KLT provides a faster algorithm than feature matching.However this comes at the cost of reduced accuracy, which is something that also holds for the choice of detectors and descriptors.