- 學位論文
基於反傳遞網路搭配雙眼立體電腦視覺做室外道路分類
Outdoor Road Classification Using Binocular Computer Vision Based on Counterpropagation Network
摘要
在本論文中，提出基於反傳遞網路搭配立體視覺去做室外道路分類。先用反傳遞網路透過色調、飽合度等資訊對影像做初步的道路分類，把影像分成道路、草地、及其它區域，再利用8鄰投票法改善反傳遞網路分類的結果，由於現在分類的結果只是局部性質相似，還要由雙攝影機所擷取的影像，計算出左右影像中特徵點的高度，在此之前，我們要先對攝影機校正，利用特定8點的3D資訊及投射到左右攝影機中之影像點來求得左右攝影機之內外部參數，之後我們利用這些參數及左右影像點進行影像中景物之3D重建。 再來是特徵點對應匹配方面，我們利用Harris Corner Detector去找出左右影像中的特徵點，再利用極線限制、fundamental matrix以及模版匹配來尋找出一組最佳的對應點。我們得知對應點的3D資訊之後，即可知道反傳遞網路分類的結果是否有誤判，以及結合反傳遞網路將影像分類的結果，把影像分成路與非路的區塊，有助於自動車導航。
並列摘要
In the study, we propose the binocular computer vision based on counterpropagation network (CPN) to classify the road outdoors. First, we employ counterpropagation network using hue and saturation for road classification initially. Second, use 8-neighbors block voting method to improve result of CPN classification, but now the classified results only have similar local properties. Third, we calculate the height of feature points in the captured image. However, we first calibrate camera. We employ the linear least square method to obtain calibration parameters of the left and the right cameras using eight known 3D points and image points projected from real world into cameras. Then we can reconstruct the 3D information by using the calibration parameters and the image points of two cameras. In the stereo correspondence, we use Harris corner detector to extract the feature points of the left and the right image. These feature points are candidates using the fundamental matrix, epipolar geometry constraints, and template matching to look for the best corresponding points. After we know 3D information of corresponding points, we can understand whether classified results of CPN mistake or not. We combine result of CPN with 3D information of corresponding points. We can divide image into road and not road area. The result is helpful to ALV navigation.