利用單眼影像進行可駕駛區域偵測與基於道路輪廓之車輛局部定位

Abstract

在自動駕駛車輛系統中，感知與車輛定位是兩個非常重要的課題。其中，感知可以提供車輛在其周遭環境的資訊，例如：其他的車輛、周圍的行人和路面與道路周遭的號誌標記。而車輛定位可以提供車輛的位置和車頭方向，並用於路徑規劃、車輛導航。藉由以上的這些資訊，車輛行駛時的安全性得以提升。 在環境感知的課題中，在車輛前方的可駕駛區域偵測是一個非常重要的研究課題。有許多不同的感測器可以用來進行道路的偵測，例如：相機、雷射、光達。而在一般在都市環境中，建構完整的道路都是由不同的色彩資訊所建立，所以基於電腦視覺的路面偵測是個很流行的課題。許多路面偵測的研究，都是利用一些機器學習的演算法來達成，像是支援向量機器和類神經網路。這些方法的準確性都基於訓練機器時的道路影像。然而，因為車輛可能遭遇的行駛環境種類繁多，要窮舉所有的駕駛環境給機器學習是件不實際的事。在本篇論文中，使用了一個名為區域延展的影像劃分演算法，其中對於道路區域的判斷依據是基於先前所定出的道路區域。藉此來判斷影像中的每個像素是否屬於道路的區域。由一開始事先定義的部分道路區塊，整個影像中的道路區塊都能延展並偵測出來。實驗結果顯示在大部分的一般車輛行駛的情境中，不論在直線行駛，遇到停止線，或是有道路標誌的情況下，找出路面的準確度都可以達到90%以上。 在車輛定位中，通常都是以衛星定位系統來取得車輛的資訊，但是在市區環境中，衛星定位系統的精確度經常會受到影響。在本論文中，提出了一個基於道路形狀的定位方法。在前一個階段，道路的區域已經被偵測出來了，而這個區域中所有的邊界都被取出，代表這個區域的形狀，並且藉由反映射投影法轉換到一個鳥瞰的座標上。藉由一個事先已知鳥瞰地圖，當車輛行駛於其中時，從拍到的影像中所擷取出的道路形狀和地圖中的道路形狀，兩者會藉由迭代最近點法來做比對，並求得車輛的位置。此外，為了提高比對的準確性，藉由比對連續時間內的影像中的道路形狀，相機的移動也可以求出並用於和地圖的比對中。而且，藉由影像資訊所求出的車輛位置也會和衛星定位系統中的位置資訊結合，相對於單靠影像資訊，結合衛星定位系統後，較差的定位結果的權重可以有效降低，求出的位置更準確。即使在直線行駛時，也能推測出車子在道路上橫向與縱向的位移。實驗結果顯示，在已規劃的道路內，藉由道路標誌、停止線、車道線定位結果中，車輛在不同時間內的絕對位置也可以被修正，而相對運動又可以和衛星定位系統十分相近。

Perception and localization are the keys in autonomous vehicle systems and driver assistance systems. The perception provides the information of environments around the vehicle, like other vehicles, pedestrians, and road signs. The localization provides the position and heading of vehicle, which can be used for path planning, navigation. With perception and localization process, the safety of vehicle driving could increase. For environment perception, the drivable road region detection ahead of the vehicle is an important research topic. Road detection can be achieved by many different sensors, including camera, laser, and LiDAR. The common structured roads in urban environment is usually constructed with different color information, so vision-based road detection is a popular topic. Many road detection researches use machine learning algorithms like supporting vector machine and neural networks whose accuracy is based on the training data. However, it is not practical to train the classifier for various scenes. In this thesis, an image segmentation method called region growing, using threshold estimated from previous indicated road region, is applied to determine that the pixels in the image belong to road region or not. With a defined initial partial road region, the whole road region can be obtained. The experimental results show that in common vehicle driving scenarios with different road markings, such as, straight road, reaching stop lines, and reaching road signs, the accuracies of road region detection results are mostly over 90%. In localization, the vehicle position is usually obtained by GNSS devices, but they are vulnerable to errors in urban scenarios. So, a contour-based localization method is proposed. The contour of road region is obtained in perception stage by extracting the boundary points or road region and transforming them into bird-eye view coordinate by inverse perspective mapping (IPM). With a prior bird-eye view map of the area where the vehicle drives, the contours of road region extracted from captured images are matching with the contour on the map by iterative closest point to obtain the vehicle position. In addition, in order to increase the precision of matching, the movements of camera are also estimated by matching the contour in consecutive frames. Furthermore, the position estimated from visual information integrated with the information from GPS to obtain more accurate position. Comparing with vision-based localization only, the integration with GPS reduces the weight and influence of bad matching results, which make the estimated position more accurate. In addition, the lateral and longitudinal movement can still be estimated even in straight road. The experimental results show that in structured road, with the localization by road signs, stop lines, and lane lines, the global positions of vehicle can be estimated while the relative movements are very close to GPS data.