2A1-E05 屋外環境における見えに基づく広域確率的位置推定(移動ロボットのための視覚)

Abstract

This paper presents a view-based localization method for vehicles and mobile robots in large-scale outdoor environments. The method works in two phases. In the first, learning phase, we acquire an image sequence while moving on a route by car. In the second, localization phase, we perform localization by comparing the input image with the learned images. An important problem in view-based outdoor localization is the change of object views due to changes of weather and seasons. Our method copes with this problem by first recognizing object by considering such view changes and then by comparing the recognition results of the learned and the input image. We use SVM (support vector machine) for object recognition and localization. We then develop a probabilistic localization method to consider the history of past movement and the uncertainty of recognition. Using a state transition model and a probabilistic model, we perform a probabilistic localization to estimate the probabilistic distribution of the car position.