利用電腦視覺與圖形識別技術在室內環境中作自動車避碰及追蹤人航行之研究

Abstract

對於許多自動化應用而言，自動車的自動航行能力節省了許多人力的消耗，而具備追蹤人航行能力的自動車更提供了許多的應用，包括用來當作自動手推車、餐車、購物車、高爾夫球車、垃圾車或除草車等等。此外，追蹤人的自動車還可以用來蒐集或學習環境的資訊，如航行路線或障礙物位置。為了達到成功導航的目的，自動車所需整合的技巧包括環境感測與學習、影像處理與特徵抽取、自動車定位與路徑規劃、以及速度與輪軸控制等。在導航的過程中，自動車必須考慮航行的安全性、平順性及穩定性，而做出不同的航行決策。在本論文研究中，我們利用電腦視覺與圖形識別技術，在室內的環境中，解決了四個自動車導航時所遭遇的問題，使得自動車能夠達到安全、平順、及穩定追蹤人航行的目的。 我們所提出的第一個方法解決自動車在室內環境航行時避免碰到障礙物的問題。這個方法是在未知的室內走廊環境中，利用圖形識別中之二次分類器找到一條避碰路徑。在這個方法中，障礙物包括走廊的牆壁以及出現在航行路線上的物體。我們利用樣本產生法將被偵測到的障礙物以及車體的兩邊分成兩類，並且當作是二次分類器中的輸入樣本。藉由分類器產生的二維判決界線，並使其通過前輪的中心，可以作為區域的避碰路徑。 第二個方法使用序列圖形識別及電腦視覺技術達到自動車追蹤人並作平順航行的目的。在自動車自動追蹤前方行走的人時，利用序列圖形識別技術來判斷行走的人是往前直走或是走得太左或太右。在這個方法中，我們拿一序列的多張影像來作決策，而自動車的方向也根據決策結果作對應的修正，以達到平順航行的目的。 第三個方法利用視覺模型及視覺接觸的限制達到自動車追蹤人時穩定航行的目的。我們提出的視覺模型包含攝影機的可視範圍以及人移動的限制範圍。這個方法的目的在於設計出一個強韌的軌跡規劃演算法，讓被追蹤的人能夠一直出現在擷取的影像中。我們發現當三個在視覺模型中的視覺接觸限制被滿足時，上述目的即可達成。第一個限制是要求人已經出現在影像中。第二個限制是要求車子下一個位置的車頭方向朝向目前人的位置。第三個限制是希望車子下一個位置和人目前位置之間的距離必須在一個限制的範圍內。根據第二個限制，我們導出了車子追蹤人時的軌跡公式。此外，我們也提出了兩個定理，證明推導出來的軌跡在實際使用上的條件。最後，我們在一軌跡規劃的演算法中，詳細描述了產生車子速度及轉角的步驟。 第四個方法是利用自動車位置的估計和使用二次分類器，達到自動車追蹤人作安全航行的目的。為了同時能夠偵測障礙物和人，我們預測人下一個可能到達的位置，並讓最接近預測位置的物體視為人，其餘偵測出來的物體都視為障礙物。再用二次分類器來找到一條追蹤人航行並作避碰的路徑。並且將被偵測到的障礙物、人位置的兩側以及車體的兩邊都當作是二次分類器中的輸入樣本。如此使得產生的避碰路徑不但能夠作為車子的航行路徑，並且能夠避免和附近的障礙物產生碰撞。 對於上述所提出來的方法，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。

Autonomous land vehicles (ALV's) are very useful in many automation applications because the ability of autonomous navigation can save a lot of manpower. Furthermore, the navigation method by which a person can lead the ALV to any desired place creates more applications of the ALV system. The system can be used as an autonomous handcart, go-cart, buffet car, shopping car, dust cart, golf cart, weeder, etc. in various applications. Besides, the system can also be used as a learning system to collect information in certain environments, including open paths or locations of obstacles. In order to achieve successful navigation, the ALV requires integration of techniques of environment sensing and learning, image processing and feature extraction, ALV location estimation and path planning, speed and wheel control, and so on. Besides, the ALV has to make decisions automatically for safe, smooth, and robust navigation. For these purposes, four approaches to obstacle avoidance and person following for vision-based autonomous land vehicle guidance in indoor environments using computer vision and pattern recognition techniques are proposed in this dissertation. In the first approach, a vision-based approach to obstacle avoidance for ALV navigation in indoor environments is proposed. The approach is based on the use of a pattern recognition scheme, the quadratic classifier, to find collision-free paths in unknown indoor corridor environments. Obstacles treated in this approach include walls and objects that appear in the way of ALV navigation in the corridor. Detected obstacles as well as the two sides of the ALV body are considered as patterns, which, after categorized into two classes, are used as input to a quadratic classifier. Finally, the two-dimensional decision boundary of the classifier, which is enforced to go through the middle point between the two front vehicle wheels, is taken as a local collision-free path. In the second approach, a new approach to ALV navigation by person following is proposed. This approach is based on sequential pattern recognition and computer vision techniques, and maintenance of smoothness for indoor navigation is the main goal. Sequential pattern recognition is used to design a classifier for making decisions about whether the person in front of the vehicle is walking straight, or is too right or too left to the vehicle. Multiple images in a sequence are used as input to the system. Corresponding adjustments of the direction of the vehicle are computed to achieve smooth navigation. In the third approach, a robust trajectory planning method for vision-based ALV guidance by person following using a visual field model is proposed. The visual field model contains a visible area and a person-bounded area. When the ALV navigates by following a walking person in front, the person has to be detected in each cycle from the image captured by a camera. The proposed trajectory planning method aims to guide the ALV to make the followed person always appear in the image. It is found in this study that if three visual contact constraints in the visual field model are satisfied, this goal can be achieved. The first constraint postulates that the person has appeared in the image. The second constraint requires that the direction of the vehicle head at the next position point straightly forward to the person's current position. The third constraint expects that the distance between the next position of the ALV and the current position of the people be bounded in a certain range. A formula for the trajectory of the vehicle that satisfies the second visual contact constraint is derived. Furthermore, two theorems specifying some conditions for the derived trajectory to be applicable to practical navigation are also derived. The steps for generating a speed and a turn angle for the ALV to conduct real-time navigation are described as a trajectory planning algorithm. Finally, the approach was tested on a real ALV. In the fourth approach, an obstacle avoidance method for use in person following for vision-based ALV guidance is proposed. This method is based on the use of vehicle location estimation and a quadratic pattern classifier, and aims to guide the ALV to follow a walking person in front by navigating along a derived collision-free path. Before generating the collision-free path, the person's location is obtained from extracted objects in the image by a person detection method. The object closest to a predicted person location is regarded as the followed person and the remaining objects are regarded as obstacles. The collision-free navigation path is designed for ALV guidance in such a way that the ALV not only can keep following the person but also can avoid collision with nearby obstacles. The navigation path results from a quadratic classifier that uses the vehicle and all of the objects, including the person, in the image as input patterns. A turn angle is then computed to drive the ALV to follow the navigation path. Our approaches are all implemented on a real ALV, and successful, safe, smooth, and robust navigation sessions confirm the feasibility of the approaches.