Localization in Urban Environments: Monocular Vision Compared to a Differential GPS Sensor

Abstract

In this paper we present a method for computing the localization of a mobile robot with reference to a learning video sequence. The robot is first guided on a path by a human, while the camera records a monocular learning sequence. Then a 3D reconstruction of the path and the environment is computed off line from the learning sequence. The 3D reconstruction is then used for computing the pose of the robot in real time (30 Hz) in autonomous navigation. Results from our localization method are compared to the ground truth measured with a differential GPS.

A Low-Cost Consistent Vehicle Localization Based on Interval Constraint Propagation

“…This property makes it possible to find correct correspondence between two images from different cameras. This feasibility has been shown by [39] in two sequences of images taken by a pair of cameras: one sequence was used in the map learning step, and the other one was used for vehicle localization; the result turns out to be robust.…”

“…The paper [39] has an average localization error of 15 cm over 6 localization experiments on 3 outdoor sequences. The reported mean indoor localization errors of [41] are ranged between 5.6 cm and 10.8 cm among 6 experiments.…”

“…We bring forward a two-stage framework (mapping and localization) to solve the problem. Closer approaches have already been proposed by different researchers [39][40][41]. They all involve a visual learning step to reconstruct a map of the environment, and they use this map for localization and navigation task.…”

“…Eric Royer et al [39] uses bundle adjustment in mapping step and the localization results are obtained via Newton iteration method. Hyon Lim et al [41] uses a structure from motion (SFM) algorithm to reconstruct the 3D environment; then in the localization step, two discrete Kalman filters are employed to estimate the camera trajectory.…”

Toward Large Scale Model Construction for Vision-Based Global Localisation

“…This method reaches real-time if the number of landmarks is quite small. Another approach is to use a full non-linear optimisation of the scene geometry: Royer (Royer et al, 2005) uses a hierarchical bundle adjustment in order to build large-scale scenes. Afterwards, Mouragnon (Mouragnon et al, 2006) proposes an incremental non-linear minimisation method in order to almost completely avoid computer memory problem by only optimizing the position of the geometry scene on the few last cameras.…”