當機器導盲犬用之有視覺自動車在人行道上導航之研究

Abstract

本研究提出了一個利用有視覺的自動車在戶外人行道上作機器導盲犬應用的系統，該系統利用一部搭載雙鏡面環場攝影機的自動車當作實驗平台，能在環場影像中直接求出實際物體的立體資訊。首先，利用環境學習的技術建立導航地圖，此地圖包含自動車導航路徑、沿途路標的位置，以及相關的導航參數。接著，利用人行道上特定的路標(人行道路緣、消防栓和電線杆)作定位來輔助導航，本研究整合上述兩項技術提出一個擁有自動定位和自動導航功能的自動車系統。 此外，本研究亦利用空間映射的方法提出新的直線偵測技術，能夠在環場影像上直接偵測出直線特徵，並計算出人行道上垂直形狀路標的位置，進而提出偵測以及定位消防栓和電線杆的方法。最後利用已定位的路標位置，來校正機械誤差，並算出正確的自動車位置。接著，本研究也提出自動跟隨人行道路緣線的技術，以及一項新的動態障礙物偵測技術，利用一「地板配對表」定出障礙物位置，讓自動車穩定且不間斷地完成導航，並在導航路徑中閃避障礙物。此外，本研究亦提出動態調整曝光值以及動態調整門檻值的技術，讓系統適應戶外環境的各種光影變化。實驗結果顯示本研究所提方法完整可行。

A vision-based autonomous vehicle system for use as a machine guide dog in outdoor sidewalk environments is proposed. A vehicle equipped with a two-mirror omni-camera system, which can compute 3D information from acquired omni-images, is used as a test bed. First, an environment learning technique is proposed to construct a navigation map, including a navigation path, along-path landmark locations, and relevant vehicle guidance parameters. Next, a vehicle navigation system with self-localization and automatic guidance capabilities using landmarks on sidewalks including curb lines, hydrants, and light poles is proposed. Based on a space-mapping technique, a new space line detection technique for use on the omni-image directly is proposed, which can compute the 3D position of a vertical space line in the shape of a sidewalk landmark. Moreover, based on the vertical space line detection technique just mentioned, hydrant and light pole detection and localization techniques are proposed. Also proposed accordingly is a method for vehicle self-localization, which can adjust an imprecise vehicle position caused by incremental mechanic errors to a correct one. In addition, for the purpose to conduct stable and continuous navigation, a curb line following technique is proposed to guide the vehicle along a sidewalk. To avoid obstacles on the navigation path, a new dynamic obstacle detection technique, which uses a ground matching table to localize an obstacle and then avoid it, is proposed. Furthermore, dynamic techniques for exposure and threshold adjustments are proposed for adapting the system’s capability to varying lighting conditions in navigation environments. Good experimental results showing the flexibility and feasibility of the proposed methods for real applications are also included.