本研究目的是要透過電腦視覺技術完成一個完整的交通標誌號與號誌偵測系統，協助駕駛人認知交通號誌與標誌，以減少駕駛人的負擔，並達到降低交通事故。 我們以實際拍攝交通號誌與標誌樣本學習偵測交通號誌與標誌色彩，以凸包演算法填補色彩空間中樣本擷取不足的色彩空缺，並以八分樹資料結構紀錄標誌與號誌的色彩範圍。透過學習到的色彩範圍作為擷取號誌與標誌色彩的依據，並將擷取出的像素連結成一個個區塊。 號誌區塊先根據長寬比、面積大小等幾何條件作初步篩選，並以號誌發光之特性作為判斷的依據。通過篩選的號誌區塊，最後以號誌特有的弧形對稱邊做確認。在偵測到的號誌周邊以邊資訊輔助偵測號誌燈箱，利用燈箱在影像中的寬度及相機內部參數推估號誌距離。 標誌區塊以幾何形狀作初步篩選，其中紅色標誌區塊通過篩選後，將剩餘的區塊正規到固定大小；經過二值化後，與事先定義好的三角形及圓形樣板比對判斷標誌形狀並分類。由於交通標誌的內容皆為白底黑字所構成，對內容物正規到固定的大小，對此區塊作疊代逼近門檻值篩選，自動找出適當的門檻值，再根據此門檻值做二值化。取得這個完整區塊的二值化影像後經由樣板比對辨識限速標誌數值。 我們在不同天候環境下測試我們的系統。交通號誌偵測的準確度為95.62%，交通標誌偵測的準確度為93.08%。限速標誌的數值辨識在772個測試資料中正確辨識755個，正確率有97.80%。 In this thesis, we propose a traffic signal and sign detection system to help drivers noticing the traffic situation on the roads. There are three stages in the proposed system: i. colored signal and sign detection, ii. signal and sign verification, and iii. signal distance estimation and sign recognition. The detection task is the most difficult due to the color is variant in different weather conditions. Here we propose a color learning method to extract the proper pixels to detect traffic signs and signals. The color distributions of traffic signal and sign are analyzed in the IHS color space. The color distribution was built by a convex hull method and described by an Octree data structure. According to the learned color distribution, candidates of signal and sign are extracted from the image. Then regions of candidates of signal and sign are verified by geometric conditions, size, and the rate of width and height. To reduce the rate of wrong detection, two special features are used to verify the detected signals. One is active light emitting, and the other is the symmetrical arc edges of a light. In this study, we only classify the signs into three classes: circle, semicircle, and triangle. To improve the practical applications, the distance between a signal and the camera is estimated based on the width of the traffic light box which can be detected by edge information around the signal region and the internal parameters of the camera. In the sign recognition, the characters of speed limit are extracted to recognize using template matching. For this purpose, we had to extract the speed limit characters. The proposed systems were evaluated in variant environments. The accuracy of traffic signal detection is 95.6% and th accuracy of traffic sign detection is 93.1%. The recognition rate of characters in speed limit signs is 97.8% from 772 character samples.
