藉由遞迴最近點之多眼相機的運動恢復結構

關於運動恢復結構這個技術，可以在許多地方發現他的應用。首先傳統在做運動恢復結構時，會使用一台單眼的照相機，一開始先在某個定點A對某3D物體拍攝，因此得到了一張照片，接著又到另一個定點B做拍攝，一樣得到另一張照片，運動恢復結構就是可以用這些不同拍攝定點所得到的2D平面照片，來去重新構築該物體的3D立體結構，這就是其中一個應用。另一個應用則是將相機移動到不同拍攝點時，兩個拍攝點之間的相機必然會存在一個轉換矩陣，我們會將相機移動某個距離，再將相機旋轉某個角度，最後在新的拍攝點做拍攝，若是我們能用電腦計算出這兩個拍攝定點的相機之轉換矩陣時，就能夠簡單又正確的移動相機的位置，進一步重建相機的移動軌跡，這樣一來可以做到數位旅遊導覽的應用，只要有大量不同拍攝點得到的2D照片，然後我們操作讓相機移動到不同的位置，就能夠得到不同位置所看到的2D照片，達到旅遊導覽的功用。本論文主要是希望能夠得到這些相機的轉換矩陣，進一步的得到相機軌跡的重建，但是我們看見目前的技術都侷限在於單眼相機來實現，我們希望透過不同相機來實現，於是我們採用了一個四眼相機，他有四個魚眼鏡頭，並且倆倆互相垂直90度，但是在使用四眼相機時，會發生傳統單眼相機沒有出現的問題，那就是當兩個相機之間的旋轉矩陣剛好在45度時，可能導致在定點A的魚眼鏡頭a能夠拍攝到某個點x，而到定點B時，變成魚眼鏡頭a無法得到該點x，如此一來在試著恢復相機的移動軌跡時，錯誤率必定會大大提升，於是我們想出另一個辦法，就是試著在每個拍攝點，透過每個拍攝點所得到的四張2D影像，將他們都擷取出特徵點之後，再把這些特徵點轉換到3D變成立體空間中的點，這樣一來，每個拍攝點，都會有他們自己的一組3D空間點雲，並且使用遞迴最近點的演算法，來將兩組3D點雲做轉換，期許能得到一個最佳的轉換矩陣，使得兩個不同拍攝點的相機軌跡重建能夠更精確。我們致力於打破傳統單眼相機的方式，透過四眼相機來重建相機的移動軌跡，希望不單單的只是像傳統那樣只能對某個區塊做重建，而是能對整個360度的場景做重建，來因應現在虛擬實境的趨勢。

關鍵字

影像處理；多眼相機；運動恢復結構；遞迴最近點；虛擬實境

並列摘要

About Structure from Motion, we can see many applications. In traditional Structure from Motion, we use single-camera to do it, in the beginning, take picture at position A, and get the 2D image, then take picture at position B, and also get the 2D image, Structure from Motion is that we use these 2D images which are taken from different positions to reconstruct the 3D structure, this is one of the application. Another application is that when we move our camera to different positions, there is a transformation matrix between two cameras of positions, if we can use computer to calculate the transformation matrix between the two position, we can easily and correctly move the camera, and then reconstruct the camera’s track, so that we can do another application such as digital tourism guide, if we have a lot of 2D images which are taken from different positions, then we can move the camera to different places so that we can see different 2D images at different positions. Out paper is that we hope to get the transformation matrix, and then reconstruct the camera’s track, but up to now we almost use single-camera to achieve it, we hope to use different kinds of camera to achieve it, so we adopt a polycamera which has four fisheye cameras, and they are perpendicular to each other, but when we use polycamera, there will be a problem that the traditional single-camera doesn’t exist, if two cameras of positions rotate almost 45 degrees, it lead to the problem that at position A, the fisheye 1 can take one point x, but at position B, the fisheye 1 may not take that point x, so that when we try to reconstruct the camera’s moving track, the error rate will increase largely, so we think out another method ,the method is that at every position, we use four 2D images at every position to calculate their feature points, and then transform these 2D feature points to 3D point cloud, so that there are 3D point clouds at every position, next we use Iterative Closest Point algorithm to calculate the transformation matrix between two 3D point clouds, we hope that we can get a best transformation matrix so as to correctly reconstruct to camera’s moving track. We devote to use nontraditional method, such as single-camera, to achieve, we use polycamera with four fisheye cameras to reconstruct the camera’s moving track, we hope that we can reconstruct not just a small zone, and we can reconstruct the 360 panaroma in response to the trend of virtual reality.

並列關鍵字

Image processing ； Polycamera ； Structure from motion ； Iterative closest point ； Virtual reality

參考文獻

[12] Google Picture.

[13] Yang Chen and Gerard Medioni, Object Modeling by Registration of Multiple Range Images, 1992

[15] Andrea Censi, An ICP variant using a point-to-line metric, 2008

[16] Johan Ekekrantz, Andrzej Pronobis, John Folkesson, Patric Jensfelt , Adaptive Iterative Closest Keypoint, 2013

[1] Wikipedia. https://en.wikipedia.org/wiki/Omnidirectional_camera

Google Scholar

國際替代計量

藉由遞迴最近點之多眼相機的運動恢復結構

主題瀏覽