路貌視差圖之分層凝聚量測法與障礙物偵測法

Abstract

本論文旨在發展可以應用於自主行動導航的立體視覺技術，目標是以立體攝像機監視路面和偵測障礙物，將路貌的三度空間資訊提供給自主機器人或自主車輛的運動控制系統，達成導航和避碰的目的。立體攝像機指向前下方，使其視野涵蓋被監視的路面，核心問題是如何從左右影像粹取監視範圍內的路貌資訊。本研究發展出分層凝聚量測法將左右影像轉換為視差圖，特點是凝聚法使量測準確、分層法則使相位估測法可以量測大範圍的視差值。路貌偵測系統由兩個機制構成，路面區域探索機制利用色調的差異區隔路面和其它物件，使物件從視差圖中浮現；視差中位數蒐尋架構則在已經去除路面區域的視差圖內，區隔出視差值大於中位數的區塊，其中區塊面積大於雜訊區塊者即被當作障礙物，最後再估算各個障礙物的三度空間位置，做為導航和避碰控制的依據，本文詳述各項技術的細節，並以實驗結果佐證其可用性。

This thesis aims at developing techniques of computational stereovision (CSV) for navigation and collision avoidance of autonomous vehicles or mobile robots. The main problem is to implement a CSV system to extract three-dimension information about roadway from stereo images. It is assumed the stereo camera looks obliquely down so that roadway is under its view. Stereo image pairs are captured and transformed sequentially into disparity maps in which roadway and obstacles are detected and located. Other than correlation correspondence method, the hierarchical coherence measurement is proposed to extract disparity maps quickly and densely from stereo image pairs. The coherence design assures precise measurement and the hierarchical design enables the phase-based disparity estimation to measure large disparities. The roadway surveillance system relies on implementing the roadway exploration mechanism and the median search architecture. The roadway exploration mechanism detects roadway pixels by the hue classification. This result is used to mask the disparity map so as to emerge objects out of the roadway. The median search architecture detects an obstacle as the region with disparity values larger than the median value and with size larger than noise patch. The regions attributed to obstacles are then located to generate three-dimension information for guidance and collision avoidance. The detailed design of the roadway surveillance system is presented. Experimental results show the feasibility and accuracy of the proposed design.