以多尺度萃取方法進行車載光達資料之類桿狀道路物件重建

Abstract

近年來隨著軟硬體技術演進，各領域如數碼城市、適地性服務(local-based services, LBS)、都市設施管理及智慧運輸系統(Intelligence Transportation System, ITS)對於細緻道路資訊需求增加，有效地獲取細緻道路資訊成為重要課題。車載光達系統(Mobile Lidar System, MLS)行駛在道路上以直接地理對位(Direct Georeference)方式進行雷射掃描，能夠快速獲取三維道路資訊以取代傳統道路調查，然而光達點雲龐大資料量及盲資料(blind data)特性，增加了車載光達應用之挑戰。 本研究目的為發展車載光達資料自動化道路類桿狀物件重建方法，道路類桿狀物件為道路之基礎設施，包含路燈、交通號誌、道路標誌、行道樹等。研究方法分為資料前處理、道路物件偵測、分類及重建。資料前處理部分，以載台軌跡為基礎進行資料切割以減低運算負擔，並濾除路廊兩側牆面；道路物件偵測部分，本文發展了一個多尺度運算架構，包含網格尺度、點尺度及重疊區處理三階段流程；物件分類以知識庫方法(knowledge-based classification)對偵測出的道路物件分類，並針對重疊區及非重疊區建立不同分類模式；最後以模型導向方式進行道路物件重建。 本研究使用的資料是以Riegl VMX-250車載光達系統獲取而得，測試地點為臺北市民權東路，總路段約為3.4公里。在偵測階段，偵測正確率約為95%，所提出之多尺度方法可較單一尺度方法提升20%以上效率，而道路物件分類部分，整體精度約為70%，主要誤差來源為受到樹冠完整覆蓋之物件，道路物件重建部分，與外業測量獲取之道路調查資料比較，位置精度約為5公分，研究成果顯示所提出的方法可有效地以車載光達資料進行三維道路物件重建。

The need of three-dimensional road modeling is gradually increasing due to the development of various applications such as cyber city, local-based services (LBS), urban infrastructure management, and intelligence transportation system (ITS). Mobile lidar system (MLS) acquires detailed and accurate 3D point clouds along road corridors. However, the blind characteristics and the huge amount of point clouds still make it difficult for application. Hence, the automatic recognition process for MLS data is needed to improve the computational time and cost for road modeling.

The objective of this research is to develop an automatic process for pole-like road objects reconstruction from MLS data. The major work includes four parts. First, the raw data is partitioned and the building façades on the roadside are removed through data pre-processing. Second, a multi-scale approach for pole detection is presented. Third, the verity of detected objects are classified through knowledge-based classification. Forth, a model-based approach is utilized for reconstruction.

The test data is acquired by Riegl VMX-250 mobile lidar system which is located at Minquan Eastern Road, Taipei, Taiwan. The length of the test area is about 3.4 kilometers. The experimental results indicates that the correctness of detection is about 95%. The overall accuracy of classification reaches 70% and the accuracy of reconstruction is about 5cm. The results indicated that the proposed method can detect and reconstruct pole-like road objects from MLS data effectively.