以電腦視覺為基礎利用特徵分群法及形狀比對技術在室外環境中作自動車之導航

Abstract

對於室內及室外環境中的許多自動化應用而言，自動車是一種非常有用的工具。成功的自動車航行所需整合的技巧包括環境感測與學習、影像處理與特徵擷取、自動車定位、路徑規畫、輪軸控制等。在室外的環境中，由於道路及道路中的物體具有潛在的複雜特性如物體中不規則及不穩定的特徵、移動的物體、陰影、損毀的路面、轉彎路、上下坡路、光線亮度的變化、甚至下雨等，所以我們需整合許多不同的演算法，甚至使用多個感測器來解決自動車導航時所面臨的種種複雜問題。在本篇論文中，我們利用電腦視覺原理，針對不同的室外道路環境，提出了五個自動車導航之方法。在傳統的導航方法中，一般而言既複雜且耗時。因此我們避免使用這些方法，而提出更有效率更快速的方法。 第一個方法我們利用彩色分群及混合循線與循路技巧使自動車能航行於寬度固定的直線道路上。該彩色分群技巧能解決自動車在航行中因光線亮度變化而產生的問題。而混合循線與循路技巧能使航行更快速、更具彈性。第二個方法我們於連續影像中，由道路中心線萃取出三條切線，而利用這三條切線來判斷自動車是否正要離開直線路段而進入轉彎路段。當自動車進入轉彎路段時，這三條切線也被用來推導此路段之航行路徑，而自動車便沿著此路徑行駛。該航行路徑假設是一個圓，而且每個航行週期重算一次以達安全之效。此外，該三條切線也可被用來判斷自動車是否正要離開轉彎路段而進入直線路段。 第三個方法允許路寬可以改變，該變化是由道路兩旁的靜止車子及車道上行駛中的車子所造成。我們避開在航行路徑上偵測車子的困難與麻煩，而只簡單地萃取出道路上可以行駛的安全路面，進而讓自動車航行於該安全路面的中心。我們使用路的邊界線來建立參考模式，並利用路面的亮度來當作視覺特徵。此外，我們提出新的演算法來求出最合適的路面形狀而不需要對路的邊界做耗時的搜尋。 第四個方法使用連續影像與座標轉換技巧來實際地偵測自動車在航行時道路前方的障礙物。當一個新的物體出現在目前航行週期的影像時，我們首先將該物體的邊緣形狀由影像中萃取出來。接著求出目前週期的自動車位置相對於下個週期的自動車位置的位移向量。然後我們利用座標轉換技巧來預測該物體的邊緣形狀在下個航行週期的影像中的位置。我們便拿此預測出來的邊緣形狀與下個航行週期的影像中所萃取出來的該物體的邊緣形狀作比對來決定該物體是否為一個障礙物。 第五個方法偵測寬度固定的上下坡路。當行駛中的車子由平坦路段進入上下坡路時，我們從影像中萃取出平坦路段與上下坡路段的左右邊界線，這些邊界線將被用來求出該上下坡路段相對於平坦路段之坡度角。當車子行駛在平坦路段而只剩上下坡路段的左右邊界線出現在影像平面上時，我們利用該坡度角來估計出平坦路段的左右邊界線，車子於是循著該估計出來的平坦路段繼續前進。在接下來的航行中，我們在航行週期的一開始便預測出平坦路段的左右邊界線，該預測出來的平坦路段的左右邊界線於是與影像中的上下坡路段的左右邊界線作比對來決定是否車子已離開平坦路段而進入上下坡路段。 對於上述所提出來之方法，我們實際地在室外環境中做了許多成功的航行測試來驗證出所提方法的有效性。

Autonomous land vehicles (ALV's) are useful for many automation applications in both indoor and outdoor environments. Successful ALV navigation requires integration of techniques of environment sensing and learning, image processing and feature extraction, ALV location, path planning, wheel control, and so on. In outdoor environments, because of the great variety of object and road conditions like irregular and unstable features on objects, moving objects, shadows, degraded regions, curved roads, ascending or descending roads, changes of illumination, and even rain, we need to combine different problem-solving algorithms and perhaps equip multiple sensors to solve the complex problem of ALV guidance in roads. In this dissertation, five approaches to vision-based ALV guidance in outdoor road environments are proposed. The conventional ways of ALV guidance, which are generally complex and time-consuming, are avoided in the proposed approaches; instead, efficient and effective ways of ALV guidance, which are usually easier and faster, are adopted. In the first approach, color information clustering and combined line and road following techniques are used for ALV guidance on straight roads with constant widths. The clustering algorithm is used to solve the problem caused by great changes of intensity in navigation. The combined line and road following technique is used to achieve faster and more flexible navigations. To locate the ALV for line following or road following, the line-model or road-model, which are constructed using path lines or road boundaries, are matched with the extracted path lines or road surface in the image, respectively. In the second approach, three tangent lines, that are extracted from the dotted central path line and collected from the images of the previous and current cycles, are used to judge whether the ALV is leaving a straight road and entering a curved road in the current cycle. When the ALV enters a curved road, the three tangent lines collected so far are used again to derive the navigation path at a curved turning road section. The navigation path is assumed to be a circle and is re-derived cyclically for safe navigation. Moreover, the three tangent lines can also be used to judge whether the ALV is leaving a curved road and entering a straight road. The third approach allows variations of road widths, which are caused by existence of static cars on the roadside or moving cars on the road lane. The conventional way of detecting obstacles and cars in the navigation route, which is in general complex and inefficient, is avoided; instead, collision-free road area detecting, which is usually easier and faster, is adopted. Road boundaries are used to construct the reference model, and the road surface intensity is selected as the visual feature in this approach. The reference model is then matched with the extracted road surface in the image to find the safe road area and the ALV location on the safe road area. In the fourth approach, image sequence and coordination transformation techniques are used to detect obstacles ahead on the safe road area in navigation. To judge whether one object newly appearing in the image of the current cycle is an obstacle, the object boundary shape is first extracted from the image. After the translation vector from the ALV location in the current cycle to that in the next cycle is estimated, the position of the boundary shape in the image of the next cycle is predicted using coordinate transformation techniques. The predicted boundary shape is then matched with the extracted boundary shape of the object in the image of the next cycle to judge whether the object is an obstacle. In the fifth approach, the ALV can keep driving forward even when ascending or descending roads appear ahead of the ALV. When the ALV keeps driving on a flat road, it detects and follows the flat road. When the ALV navigates at a transition from a flat road into an ascending road, both of the flat and the ascending road boundaries are extracted from the image, which are then used to estimate the slant angle of the ascending road and compute accordingly the connection points of the flat and the ascending road boundaries. When the ALV drives on the flat road but the flat road boundaries disappear from the image, the flat road boundaries are derived using the estimated slant angle, the stable features of the ascending road boundaries in the image, and the predicted connection points of the ascending and the flat road boundaries. The ALV then follows the derived flat road because it still drives on the flat road. At the beginning of subsequent each navigation cycle, the ALV predicts the flat road boundaries in the image using the derived flat road boundaries just described in the previous cycle. The predicted flat road boundaries are then matched with the extracted ascending road boundaries in the image to judge whether the ALV has entered the ascending road in the current cycle. This way of guidance is also used when the ALV navigates at a transition from a flat road into a descending road. A lot of successful navigation tests show that the proposed approaches are effective for ALV guidance in outdoor road environments.