基於近紅外線廣角相機的自我定位演算法之研究

Abstract

即時定位與地圖構建(SLAM)是一種用來解決機器人自我定位問題的通用方法。目前主流的SLAM方法主要分為直接法和基於特徵點的方法。直接法對於亮度資訊比較敏感，基於特徵點的方法對於亮度變化有一定的容忍度。在夜間環境下，正常相機的信息變得很難識別。因此，我們提出了使用近紅外攝影機進行夜間機器人的自我定位。通過近紅外線(NIR)得到的影像會因為距離紅外線燈的遠近而出現不同的光照強度的變化，所以不適合用直接法進行SLAM。我們使用了對光照有一定容忍性的基於特徵點的方法進行夜間自我定位。該方法將結合多台廣角红外相機以及近紅外線進行自我定位，不僅可以解決低亮度條件下特徵點的捕捉問題，還可以獲得準確的定位結果。使得機器人在室内夜間低亮度的情況下仍然可以繼續進行自我定位，從而更好地適應複雜多變的情況。

Simultaneous Localization and Mapping(SLAM)is a generic method used to solve robot ego-positioning problems. At present, popular SLAM methods are mainly divided into direct methods and feature-based methods. The direct method is more sensitive to brightness information, and the feature-based method has been demonstrated more tolerance toward changes in brightness. In nighttime environment, information from normal camera become difficult to identify. Therefore, we propose to use near-infrared (NIR) cameras for ego-positioning of night robots. Images obtained by NIR light have different light intensities due to their distance from the infrared light sources, so SLAM is not suitable for direct methods. We used a feature-based method which is tolerant to light intensities to ego-positioning. This method combines ego-positioning and NIR with multi-wide-angle NIR cameras to not only solve the problem of capturing the feature points under low-light conditions, but also obtain more accurate ego-positioning results. This allows the robot to do SLAM even at night in low illumination indoor scenario. Therefore, this system has more robust results in different challenging situations.