利用電腦視覺技術做自動車航行之室內環境學習及導航

Abstract

對於自動車的導航，我們經常採取模組比對的方式。然而，建立環境模組是一件麻煩又耗時的工作，所以我們需要一套可以自動產生環境模組的方法。如果環境模組可以經自動學習而產生，自動車系統的可變度便可得到提升。此外，學習得到的環境模組也可做為虛擬實境或其他應用的場景建構之用。在這篇論文中，我一共針對室內自動車航行之環境學習及導航提出了四套不同的方法。 第一個方法是一種漸進式的學習方法。這個方法包括了三個階段：初始學習，航行以及更新環境模組。在初始學習階段中，自動車由人駕駛，而在此同時，航行環境的影像以及其他資訊同時被自動記錄下來。爾後，經由一離線的學習程序後便可得到初始的環境模組。在航行階段中，自動車可自行在學習過環境中航行，一方面用模組比對的方式來定位，另一方面同時記錄下更新環境模組所需要的資訊。在更新環境模組階段，我們利用一離線程序來更新學好的環境模組。在每次經過這樣的『航行而後更新』的程序後，我們可以得到比原來更精確的環境模組。在本方法中所操用的環境特徵為影像中的垂直線。此外，我們提出了多權值的廣義赫式轉換來解決模組比對的問題。 第二個方法是一套智慧型的環境模組及導航策略學習系統。在此方法中，自動車先由人手動操作走過其航行環境，爾後，此學習系統便可自動產行一套針對此環境的導航模組。導航模組所含的資訊包括環境特徵的位置、經修正後的航行路徑以及在航行此路徑時所需採取的導航策略。導航策略的選擇由當時可以偵測得到的穩定環境特徵的數量及種類來決定。此外，我們也提出了一套可以綜合各種導航策略的航行機制，用以應付多變的航行環境。此航行機制並可利用學習而得的導航模組來做安全導航。 第三個方法是利用下推式推理機來做無人干涉的全自動環境學習。在此方法中，我們設計出一個下推式推理機來幫助自動車探索完整個航行環境。經感測量得的環境特徵經過前處理器編碼後成為下推式推理機的輸入資訊，而代表某特定自動車動作的下推式推理機輸出可以用來指揮自動車的動作。此外，在此還提出了對我們所提出的下推式推理機之的理論性正確證明，並且做出一套模擬系統，證明我們所提出的下推式推理機可以正確的工作。 除此之外，我們還將第三個方法所提出的理論實做在真實環境的學習上。在此方法的真實環境實作中，自動車可以不經任何人為的操作，有系統地自動航行於一未知的封閉區域，並同時記錄模環特徵以及其他資訊，最後自動建立出航行環境的俯視地圖以為航行或其他應用之用。此學習系統包含三個子系統：環境特徵定位子系統，環境模組管理子系統以及環境探索子系統。環境特徵定位子系統處理輸入的影像，並且利用模式比對的方法得到環境特徵以及自動車本身的所在位置。為了簡化環境特徵的偵測，我們在自動車上加裝了兩部平面雷射。雷射光在牆面上的投影很容易被偵測得到，由此就可以較容易得知牆面的位置。環境模組管理子系統利用比對轉角的位置，將所得到的區域環境特徵加到全域環境模組中。環境探索子系統利用下推式推理機、目前已學得之環境模組以及自動車目前的位置導引自動車探索完整個航行環境。 對於上述所提出來的幾個方法，我們在真實的自動車上實際執行，並做了許多成功的航行測試，驗證了所提方法的可行性。

Model-based guidance approaches are often employed in autonomous land vehicle navigation. However, the establishment of environment models is a time-consuming work. It is thus desired to design a process for automatic modeling of navigation envi-ronment. With automatic environment learning, the flexibility of certain ALV navigation systems is increased. Besides, the learned environment models can also be used for vir-tual reality scene construction and other applications. In this dissertation, four ap-proaches to environment learning for vision-based ALV navigation in indoor environ-ments are proposed. The first is an incremental-learning-by-navigation approach. The approach consists of three stages: initial learning, navigation, and model updating. In the initial learning stage, the ALV is driven manually, and environment images and other status data are recorded automatically. Then, an off-line procedure is performed to build an initial en-vironment model. In the navigation stage, the ALV moves along the learned environ-ment automatically, locates itself by model matching, and records necessary information for model updating. In the model updating stage, an off-line procedure is performed to refine the learned model. A more precise model is obtained after each naviga-tion-and-update iteration. Used environment features are vertical straight lines in cam-era views. A multi-weighted generalized Hough transform is proposed for model matching. In the second approach, an intelligent system for learning environment models and guidance strategies for vision-based ALV navigation in indoor environments is proposed. In the learning process, the ALV is firstly driven manually by an operator through the navigation environment. Then, a navigation model, which consists of the locations of environment features, refined sub-paths, and their corresponding guidance strategies, is generated automatically by the learning system. The selection of the guidance strategy depends on the availability of stable environment features. An intelligent navigation scheme by integrating three guidance strategies is proposed for safe ALV navigation through environments consisting of various conditions. The learned model can be used to guide the ALV through the explored environment by the proposed navigation scheme. The third proposed approach is unsupervised learning by the use of an automata model called pushdown transducer. In the learning system, a pushdown transducer is employed to guide the ALV to explore the entire navigation environment. The sensed local environment features are encoded into symbols for use as input into the pushdown transducer by a preprocessing unit, and output symbols representing special ALV ac-tions are generated by the transducer to guide the ALV. Some theoretical proofs showing that the proposed transducer is able to explore the entire region of any closed navigation environment in finite cycles are included. A learning system for simulated grid model environments was implemented and several simulation results show the feasibility of the proposed approach. For the third approach, we have accomplished a practical implementation for use in real indoor environments. In its implementation, the ALV may, without human’s involvement, self-navigate systematically in an unexplored closed environment, collect the information of the environment features, and then build a top-view map of the envi-ronment for later planned navigation or other applications. The learning system consists of three subsystems: a feature location subsystem, a model management subsystem, and an environment exploration subsystem. The feature location subsystem processes input images, and calculates the locations of the local features and the ALV by model match-ing techniques. To facilitate feature collection, two laser markers are mounted on the vehicle which project laser light on the corridor walls to form easily-detectable line and corner features. The model management subsystem attaches the local model into a global one by merging matched corner pairs as well as line segment pairs. The envi-ronment exploration subsystem guides the ALV to explore the entire navigation envi-ronment by using the information of the learned model and the current ALV location. The proposed approaches have been tested on a prototype ALV and many suc-cessful navigation sessions have been performed, which confirm the feasibility of the proposed approaches.