Mobile robot outdoor localization using planar beacons and visual improved odometry

Abstract

This paper presents experimental results on the localization of a mobile robot equipped with relative frequent and absolute infrequent sensors. The relative sensors used are two: a wheel based odometry and a visual based odometry. The absolute sensor is a vision based landmark detector that computes the pose of the robot relative to a pre-mapped visual beacon. This would be a simple sensor fusion problem, which could be solved using standard recursive estimators, if we would not have considered two extra characteristics of the beacon detector: (1) since we assume a monocular vision system and a planar visual mark, the localization problem presents up to four possible solutions; and (2) the frequency that the robot meets a visual mark is very low (0.01Hz or less). To consider these characteristics, we propose the use of a particle filter with a very precise prediction step (obtained by combining the two odometry sensors available) and a correction step that considers the multi-modal characteristic of the data. Besides, the sensor fusion algorithm, the paper also describes the development of the visual sensors used in the localization process.

Omnivision-based KLD-Monte Carlo Localization

“…In [18], a localization system using a particle filter and a monocular camera was presented. The Scale-Invariant Feature Transform (SIFT) signature of the images in a database was used for comparison with the present image.…”

A Trinocular Stereo System for Detection of Thin Horizontal Structures

“…The algorithm approaches near real-time. In [11] they combine visual odometry with wheel based odometry to get a relative position. SIFT is used for landmark detection to get an absolute position.…”

Improving Odometric Accuracy for an Autonomous Electric Cart

“…For example, GPS loses coverage when the vehicle does not have a full sky vision [ 9 ]. As another examples, if we introduce beacons, we need to structure the whole environment, or a laser needs available features to recognize in the environment, and these features should be in the range of action of the laser (typically between 10–20 m) [ 10 ].…”