基於視覺建構自動導航車的智慧型環境景物辨識

Vision-Based Intelligent Scenery Object Recognition Applied to Autonomous Land Vehicle

賴錦德(Ching-Te Lai)

指導教授 ： 駱榮欽

國立臺北科技大學/電資學院/電腦與通訊研究所/碩士(2010年)

摘要

本篇論文中，我們提出一種視覺自動導航車於行車時，前方環境景物的辨識系統，用來輔助室外自動導航的路徑規劃與定位。此系統使用改良式的光流法從連續影像中取得移動的影像區塊設為動態候選區塊，並依據光流子的特性來取得這些區塊的行為及移動速度等資訊。同時以視差圖方法從水平左右影像中求得影像中各區塊的高度與寬度等資訊。之後再根據自動導航車的位置和行進路徑等資訊，將影像上影響行車安全的區域，設為感興趣的區域。而後使用類神經網路的多分類器及樣板比對法來辨識各區塊內的影像。最後若確認為障礙物並影響行車安全，則依設定的因應策略來做警示、閃避或忽視等動作。
在系統架構上是透過兩台水平的CCD攝影機，藉由影像擷取卡將左右影像同時輸入車載桌上型電腦。透過影像前處理和光流場及視差比對等運算，最後分析出各種類型的影像。以物件導向的概念將其設為物件，並建立類別來詳述針對這些物件的屬性、行為及方法。此概念轉換到實際上，就是針對各種偵測到的障礙物或非障礙物，根據其特性，做相對應的策略。由於行車影像的背景與環境景物會持續變化，因此使用改良式的光流法來去除相對移動的向量，同時取得障礙物的方向與移動速度；再配合視差圖取得高度與寬度的資訊以判斷是否為需要閃避的障礙物；使用多分類器和樣板比對法可以成功辨識出各類型的影像。最後完成整個環境影像系統的建構。

關鍵字

自動導航車 ； 光流場 ； 視差圖 ； 類神經網路 ； 多分類器 ； 樣本比對

並列摘要

In the paper, we propose a vision-based scenery object recognition system while the autonomous land vehicle (ALV) is navigating on a road, in order to support its path planning and locating. In the system, we use improved optical flow algorithm to extract the blocks of moving objects in sequential images, and we set them as the moving candidate blocks. By way of optical flow computing, we can find the moving behaviors and velocities of the blocks. At the same time, by the disparity map (DM), we obtain the height and width information of the blocks from a stereo pair of images. Then, according to the information of location and moving path of ALV, we can identify an area as a region of interest (ROI) in the image for the safety of navigation. In the study, we use multiple classifiers of neural network and template matching method. By way of matching with prior databases and multiple classifiers, we could recognize the attributes of all the detected blocks. Finally, if we verify that an obstacle does affect the safety of navigation, we set strategies to alarm the system to dodge it or to ignore it.
The system architecture includes two horizontally parallel CCD video cameras in front of ALV and one desktop PC with capture cards in ALV. We obtain both sides of sequential images in the PC through the capture cards. After image preprocessing and computing by optical flow and DM, we analyze several types of image data. By the principle of an object oriented design (OOD), we set them as objects, and also create classes to describe the attributes, behaviors, and methods of them. Transforming it into actuality is like how ALV controls itself to the detected obstacles or nun-obstacles. We choose an improved optical flow algorithm to fit on the changing background of images, and we can also obtain moving behaviors and velocities of objects. By DM, we have height and width information of obstacles to set ALV’s navigational strategies. After computing by multiple classifiers and template matching, the system could recognize the attributes and the behaviors of the objects from the scenery images in ALV navigational environments.

並列關鍵字

ALV ； Optical Flow ； Disparity Map ； Neural Network ； Multiple Classifier ； Template Matching

參考文獻

[2] J. Shin, S. Kim, S. Kang, S. Lee, J. Paik, B. Abidi, and M. Abidi, “Optical Flow-based Real-time Object Tracking Using Non-prior Training Active Feature Mode,” ELSEVIER Real-Time Imaging, Vol. 11, June 2005, pp. 204-218.

[3] J, Ma and N, Ahuja, “Region Correspondence by Global Configuration Matching and Progressive Delaunay Triangulation,” IEEE Conference, Vol. 2, June 2000, pp. 13-15.

[4] S. Birchfield and C. Tomasi, “Depth Discontinuities by Pixel-to-pixel Stereo,” Proceeding IEEE International Conference on Computer Vision, 1998, pp. 1073-1080.

[5] Z. Li and X. Tang, “Using Support Vector Machines to Enhance The Performance of Bayesian Face Recognition,” IEEE Transactions on Information Forensics and Security, Vol. 2, No. 2, June 2007, pp. 174-180.

[6] Limin Xia, “Vehicle Recognition Using Boosting Neural Network Classifiers,” Proceedings of the 6th World Congress on Intelligent Control and Automation, June 2006, Dalian, China, pp. 9641-9644.