隨著電腦科學的進步與社會安全的需要，利用影像處理的技術針對空照影像中地面上感興趣的物件進行識別成為近年來熱門的研究主題。這樣的技術運用在軍事上，可以對敵方進行戰場情報蒐集，應用在天然災害防治上，可以深入深山叢林中，協助尋找等待救援的人員。應用在交通運輸中，可以發展成空中智慧型監控系統，進而提供警察執行治安與道路秩序維護之協助工具，不僅降低人為誤失更節省大量人力資源。對於空中智慧型監控系統而言，正確的車輛偵測更是扮演重要的角色。
在本研究中，我們提出一套適用於空照影像的車輛自動偵測系統，而此系統可分為三個大部分：第一部份是背景色彩的濾除，由於空照影像中車輛在整個畫面的中所佔的比例相當少，因此將絕大多數的背景資訊濾除，可以加快後面步驟的處理速度。第二部份是車輛特徵擷取，此系統中我們所使用的特徵包括車輛色彩、邊緣資訊以及局部特徵點資訊。在車輛色彩擷取的部份，由於車子本身是屬於金屬材質，在光線的反射上會與一般非金屬材質的物體有所不同，透過色彩空間轉換可以將車輛色彩更有效的區別出來。在邊緣資訊的擷取，我們加入了衝量保持法使得邊緣偵測器可動態調整閥值，而適應於各種不同的空照影像。第三個部份是動態貝氏網路（Dynamic Bayesian Network）結構的建立與訓練，利用訓練好的網路對第二部份得到的車輛特徵作辨識，判斷像素點是否屬於車子的一部分。
實驗部分採用了數段不同的空照影像，在未經過訓練的測試影像中平均車輛偵測準確率與錯誤偵測個數分別是92.04%與0.16616，實驗結果顯示論文所提出的方法能在各種空照影像中正確且有效的偵測車輛。

With the advancement of computer technology and increasing needs of social security, studies of target object detection in aerial surveillance using image processing techniques are growing more and more important. These technologies can be employed in various applications, such as gathering enemy information for military purpose and searching for missing people in mountain areas. In intelligence transportation applications, aerial surveillance not only provides traffic monitoring but also assists traffic management, which can save a lot amount of human resources. Among the basic modules in intelligent aerial surveillance, vehicle detection plays a very important role.
In this thesis, we present an automatic vehicle detection system for aerial surveillance. First of all, background colors are eliminated and then features are extracted. In this system, we consider features including vehicle colors, edges and local feature points. For vehicle color extraction, we utilize color transform to separate vehicle colors and non-vehicle colors effectively. For edge detection, we apply moment-preserving method to adjust the thresholds for canny edge detector automatically, which increases the adaptability and accuracy for detection in various aerial images. Afterwards, a Dynamic Bayesian Network (DBN) is constructed for classification purpose. Based on the features extracted, a well trained DBN can estimate the probability of a pixel belonging to a vehicle or not.
In this work, features are obtained from a neighborhood region but the detection task is based on pixel-wise classification, which is more effective and efficient than multi-scale sliding window or region-based methods. Experiments were conducted on a variety of aerial videos and the average hit rate of testing videos is 92.04%. The number of false positives per frame of testing videos is 0.16616. The experimental results reveal that the proposed approach is feasible for generalization and effective for vehicle detection in various aerial videos.

[1]S.M. Al-Garni and A. A. Abdennour, “Moving Vehicle Detection using Automatic Background Extractuib,” World Academy of Science, Engineering and Technology 24,2006
[2]R. Lin, X. Cao, Y. Xu, C. Wu, and H. Qiao, “Airborne moving vehicle detection for video surveillance of urban traffic,” Intelligent Vehicles Symposium, 2009
[3]L.W. Tsai, J.W. Hsieh, Member, IEEE, and K.C. Fan, Member, IEEE, “Vehicle detection using normalized color and edge map,” IEEE transactions on image processing, VOL. 16, NO. 3, 2007
[4]J.Y. Choi and Y.K. Yang, “Vehicle detection from aerial images using local shape information,” PSIVT 2009, LNCS 5414, pp.227-236, 2009
[5]H. Cheng and D. Butler, “Segmentation of aerial surveillance video using a mixture of experts,” IEEE Digital Imaging Computing: Techniques and Applications, 2005
[6]C. Tzomakas and W. Seelen, “Vehicle detection in traffic scenes using shadow,” Tech. Tep. 98-06 Inst. fur neuroinformatik, Ruhtuniversitat, Gemany, 1998
[7]C.G. Harris and M.J. Stephens, “A combined corner and edge detector,” Proc. 4th Alvey Veision Conf., Manchester, p.147-151, 1988
[8]L. Wang and Y. Shen, “An Algorithm Based On the Harris Corner Detector to Identify the Palmprint Extraction,” ICEMI, 2009
[9]J.F.Canny, “A Computational Approach to Edge Detection,” IEEE Trans Pattern Analysis and Machine Intelligence, 8(6):p.679-698, 1986
[10]C. Cortes and V. Vapnik, “Support-Vector Networks,” Machine Learning, VOL. 20, NO. 3, p.273-297, 1995
[11]H. Simon, “Neural Networks: A Comprehensive Foundation,” 1998
[12]S. Rusell and P. Norvig, “Artificial Intelligence : A Modern Approach,” p.712-731
[13]N. Friedman, M. Goldszmidt, D. Heckerman, and S. Russell, “Challenge: Where is the impact of Bayesian network in learning?”, proceedings of the Fifteenth International Joint Conference on Artificial Intelligence, Morgan Kaufmann Publishers, 1997
[14]Z. Ghahramani, “Learning Dynamic Bayesian Networks”
[15]C.K. Yang and W.H. Tsai, “Reduction of color space dimensionality by moment-perserving thresholding and its application for edge detection in color image,” Pattern Recognition Letters 17, 481-490, 1996
[16]I.P. Chen, M.Y. Lai and L.L. Wang, “An improvement in the moment-preserving thresholding method,” International Journal of Imaging Systems and Technology, vol. 18, p.365-370, 2008