精準且有自適能力之環場視覺 技術及應用之研究

Abstract

為了能使電腦與四周環境互動，環場視覺是一項極其有效且十分重要的技術。與傳統電腦視覺技術相比，環場視覺強調其在單一時間點能對大範圍環境取景之能力，而不用在攝影機上加裝馬達裝置來週期性地轉移攝影機，更不需用多部攝影機來對環境取景。由上述可知，在環場視覺技術中我們可以避免影像接合、攝影機換手、多攝影機特徵連續追蹤等複雜問題。為了達到大範圍取景之目的，有兩種特殊設計的環場攝影機較常被使用，其一是反射式環場攝影機，另一種是魚眼攝影機。其中，前者是將一個特殊形狀的反射式鏡面擺放在一傳統攝影機前方，藉由該鏡面來研展攝影機的可視範圍；後者是利用一特殊的魚眼透鏡來研展其可視範圍。然而，因為極大範圍的環境資訊被濃縮於一張傳統大小的影像中，環場攝影機所擷取到的影像必定有十分嚴重的扭曲，這也使得後續影像分析的工作變得困難且複雜許多。雖然將影像扭曲校正回來是其中一種簡單的解決方式，然而因為扭曲造成影像解析度的不同，校正回來的影像在某些區域會十分地模糊，在影像分析後會產生不穩定的結果。更甚者，上述扭曲校正的過程也需要些許運算能力，在即時應用及嵌入式系統中都較不適用。 為了克服環場攝影機擷取到的嚴重扭曲，我們提出了在扭曲影像上精準且穩定地偵測空間中直線的方法。另外，我們也針對各種利用環場攝影機偵測空間中直線的應用，提出改良的攝影機模型，並且也提出一套方便的校正程序來校正環場攝影機。此校正程序只需使用空間中的一條直線特徵，且不需要測量其位置及方向，使得整個校正程序變得十分簡單，且可讓一般使用者方便地進行校正，使環場視覺技術朝消費電子更邁進一步。 另外，從消費者的角度上來看，我們應該也要能讓一般使用者方便地架設一套環場視覺系統。在此方面，我們提出一套新的雙眼視覺系統，此系統可讓使用者任意地擺放兩部環場攝影機。在擺放完成後，系統會自動利用環境中之直線特徵來回推攝影機的位置及角度，從而正確地計算立體資訊，以供各種人機互動應用使用。另一方面，針對需要取得十分精準之立體資訊的應用中，我們也提出一套最佳化架構以及三個最佳化演算法，其可告訴使用者如何擺放該二部環場攝影機的位置及角度，以求得最佳之立體資訊。根據這些最佳化演算法，使用者將可建構出能進行精準立體測量之雙眼環場視覺系統。 最後，我們也對上述所提出之各種環場視覺技術進行延伸研究，開發一套室內停車場管控系統。此系統可利用假設於天花板之各環場攝影機，自動地分析停車場中各停車格之位置，並自動找出空的停車格位置，以利駕駛找尋停車位。與現有系統相比，我們提出的系統因為攝影機的可視範圍較大，所以只需要較少的攝影機數量；另外，因為我們提出的系統可自動分析停車格位置，因此其系統建置過程會便利許多。 在可行性及效率評估中，我們已對上述各方法及技術進行理論及實驗分析，並得到良好之實驗結果。

Omni-vision is an important and effective technique to make computers be aware of the surrounded environment. Different from traditional computer vision techniques, omni-vision ones emphasize more on capturing the environment information within a very wide area at one time without adding a motor control to the camera, moving the camera periodically, or using multiple cameras. Such techniques can avoid the difficulties of image stitching, camera hand-off, feature tracking over different cameras, etc. To achieve the capability of capturing information of a wide area, two special kinds of cameras are commonly used, which are catadioptric omni-directional cameras, and fisheye-lens cameras. The formal ones use a specially-designed reflective mirror to extend the viewing field, and the latter ones use a fisheye-lens to achieve the goal. However, since the environment information captured from a wide area is all compressed in a relatively small image, the captured image is inevitably heavily distorted, which makes the image analysis task much more difficult and complicated. Although, an easy and feasible way to deal with the heavy distortion is to unwarp the captured images to yield an image looking like one captured by a conventional perspective camera. However, since the resolution distributions captured by omni-directional cameras and by conventional perspective camera are quite different from each other, an unwrapped image becomes much more blurred in some regions, making image analysis tasks unstable and unreliable. Furthermore, the unwarping process needs some computation power, making it unsuitable to real-time applications and embedded systems with restricted computation power. To deal with the heavily-distorted images captured by omni-directional cameras, an accurate and reliable space line detection method without unwarping the distorted image is proposed. Also, to model the imaging process conducted by an omni-directional camera, a new camera model along with a convenient calibration process to calibrate an omni-camera easily is proposed. This new calibration technique requires only one straight line in the environment without knowing the position or direction of the line, making it possible for non-technical user to conduct the calibration work which is a big step toward consumer electronics. In addition, from the viewpoint of a consumer, the setup procedure of an omni-vision system should be sufficiently convenient for use by a typical user with no technical background. In this sense, a new binocular omni-vision system is proposed, which allows the user to place the two omni-directional cameras freely at any positions and with any orientations. After the two cameras are placed, the system can automatically derive the cameras’ positions and orientations via analysis of the space lines within the environment. As a result, the binocular omni-vision system can calculate 3D information correctly for use in many advanced human-machine interaction applications. Furthermore, for applications requiring precise 3D information, an optimization framework along with three different optimization algorithms are proposed as well to tell the user where to place the two omni-cameras, and what are the best orientations. With these optimization algorithms, the user can set up a binocular omni-vision system which acquires the most precise 3D data. Finally, the proposed omni-vision techniques are extended for uses in the application of indoor parking lot management. The proposed system for this application utilizes the omni-directional cameras mounted on the ceiling, and automatically analyzes the acquired images to obtain the locations of the parking spaces and detect vacant parking spaces. Different from existing similar application systems, the proposed one requires fewer cameras due to the wider fields of view of the cameras, and is much more convenient to set up because of the developed automatic parking-space analysis capability. The feasibility and effectiveness of all the above proposed methods and systems are demonstrated by theoretical analyses and good experimental results.