具動態三維投影模型之環車視訊系統

Abstract

隨著人們對於行車安全的重視，有關安全輔助的車輛電子產品越來越多元，像是汽車防撞、倒車雷達、倒車顯影等，這些產品提供了給使駕駛者直觀的影像，好讓駕駛者了解車輛四周的狀況。而我們研發的車輛環周系統，將還原最真實的景象，有助於駕駛者釐清盲點。

市面上已經出現了一些車輛環周系統，主要透過四台相機，拼接成一張全景圖，供駕駛者釐清車周的盲點。但是在影像重疊的區域仍是眾多難題所在，最主要的解決方法有兩種：一種是選擇一條適當的接縫，將兩張相鄰影像一分為二，但此種作法可能會讓在接縫處的物體消失；另一作法是將重複區域的像素加權平均，此方法的缺點是會產生重影問題，而在此研究中我們採用的是後者。因此我們將著重在解決影像接合的重影，透過3D模型的變形，使得接合後的影像更真實。

在過去，三維全景圖將影像貼在球面或是圓柱面上，但是在我們的環境上，相機視差過大以致於會有鬼影產生，因此我們透過特徵點匹配，找出物件在真實世界的座標，將投影模型做變形，使得此模型盡量符合真實世界的形狀。此研究將著重於大視差攝影機之特徵點匹配與投影模型之形變，而採用了形變的技術後，就能達到消除重影的目的。

我們採用相機校正、投影模型之形變及顏色調整的技術於系統上。在實作方面，我們架設四支魚眼攝影機，並採用基於特徵點的形變於投影模型，透過相鄰攝影機之特徵點匹配來消除重複區域之重影。

Nowadays, more and more automobile electronics appears since people attach great importance to safety issue. For example, collision avoidance, parking sensor, blind spot monitor. Those devices provide intuitive image information to driver and let driver to see the view around his vehicle. In this work, we developed a surround view monitoring system to provide a realistic panorama which helps driver to aware of blind spots.

There are some omniview products commercially available. They used four cameras to build a panorama, but it was hard to fuse the overlapping region between two camera views. There are two approaches to deal with overlapping region. One approach is to find an appropriate seam and divide images into two parts. The disadvantage is that some objects in the overlapping region may be missing. Another approach is image blending by summing weighted pixels, but it may have ghost effect. In order to avoid those two defects, we applied blending method in our research and eliminated ghost effect by deforming 3D projection model. Then we can make the panorama more realistic.

In the past, 3D panorama was created by mapping texture images on a sphere or cylinder. The parallax between cameras is large in our environment, so ghost effect appears when cameras are not at the same projection center. We used feature matching to obtain the position of objects, and deform the projection model to fit objects. The main contribution of our work is feature matching on large parallax and deformation on the projection model. By using deformation technique, the proposed system could eliminate ghost effect in the overlapping region.

In our research, we use camera calibration, feature matching on large parallax, projection model deformation and image blending to construct our system. In the implementation, we mounted four fish-eye cameras to capture images. By using feature-based deformation on the projection model, we could eliminate ghost effect in the overlapping region with matching features detected in neighboring camera views.