基於幾何圖形分析之交通影像處理演算法

Abstract

駕駛人輔助系統為近年來車輛上不可或缺的基本配備，更為自動駕駛汽車中重要的裝備。自動交通標誌的偵測與辨識為可以優化駕駛人輔助系統的科技，使駕駛人可以注意到各個路段的警告、禁制或是指示等資訊。另一方面，估測前方車輛距離的功能也是駕駛人輔助系統中重要的一環。 世界各國的交通標誌皆有自己的既定格式，但各國交通標誌的設計有一些特徵使人們可以簡單的識別。以台灣的交通標誌為例，特定的形狀以及顏色分別代表不同層級的警告標示，且交通標誌放置在路面特定高度及位置，因此在車輛上裝置的鏡頭更可以預測交通標誌的旋轉以及幾何變化。 本篇論文中我們收集了台灣的10組不同的交通標誌，每組分別有30-40張不同街景、不同明暗度、不同天氣狀況下所拍出來的照片。我們所提出來的交通標誌辨識系統，總共分為五個步驟，第一個步驟為將圖片進行前處理，將輸入的影像縮小至固定的大小以減少運算量，且將色彩空間從RGB轉換至HSV。第二步驟為偵測交通標誌，將處理完的圖片先使用快速的色彩偵測過濾掉背景，再使用霍夫轉換以及多邊形近似演算法等形狀辨識方式偵測出我們欲辨識的交通標誌。第三步驟將第二步所偵測出的區域切割出來後使用哈里斯邊角偵測提取特徵，將提取的特徵點利用仿射轉換、單應性、主成分分析等方式將圖片與標準圖片做對齊。第四步驟將對齊完成的圖片根據圖片的明暗度做二值化。最後一個步驟與標準圖二值化後的像素做比對，進而辨識出交通標誌。 前方車輛距離估測的演算法中，我們分別取了五種不同的車輛距離的影像。使用影像中前方車輛的輪胎與路面的接觸點以及自己車上的單鏡頭相機之間的幾何關係計算出前方車距再使用路面標線做校正。我們的演算法得出的平均相對誤差率為1.0546%, 較其他使用單鏡頭估計前方車距的方式準確。 交通標誌偵測以及識別演算法中，在資料量不足的狀況下亦能達到非常高的準確率，在不同環境中例如天色較暗或是雨天所拍攝的圖片皆可有效辨識。且在一般沒有GPU的電腦中也可以快速有效的辨識交通標誌，在我們的資料庫中達到100% 的辨識率。

Nowadays, the driver assistance system is the basic device on the vehicle. The driver assistance system is also essential for the self-driving vehicle. The automatic traffic sign detection and recognition system can optimize the driver assistance system through reminding drivers to pay attention to the warning, prohibition, or instruction traffic signs on the road. On the other hand, the front vehicle distance estimation system is also important for the driver assistance system. It can announce the driver about the distance of the car in front of it. Taking traffic signs in Taiwan as an example, specific colors, shapes, and sizes of traffic signs concern different levels of warning. We can use the camera on the vehicle to capture the street image and identify traffic signs with the predictable rotation and geometric change easily. In this thesis, for the traffic sign detection and recognition part, we collected 10 different classes of traffic signs and each class contains 30-40 images under different street scenes, brightness, and weather conditions. We proposed a traffic sign detection and recognition system. The system is divided into five steps. The first step is pre-processing, including resizing the image into smaller scale and transforming the color space from RGB to HSV. The second step is traffic signs detection. We use color detection based on the HSV color space to eliminate background regions, and then use the Hough transform and the approximate polygon algorithm to detect the shape of traffic signs. The third step is alignment. We crop the region extracted from the second step and use Harris corner detection to find feature points. Furthermore, we apply several techniques, including affine transformation, homography, PCA to align the input images with standard images. After alignment, binarization is performed. The final step is comparing input images and standard images pixels by pixels to identify the class of the input image. For the distance estimation part, we calculate the distance of several images. There are 5 different vehicle distances in our dataset. We estimate the distance by the geometric relationship between the monocular camera and some feature points in the image. The average error rate of the distance estimation is 1.0546%, lower than other state of art methods based on monocular vision. The proposed methods of traffic sign detection and recognition system can achieve high accuracy even when the amount of data is insufficient. We can also recognize the traffic signs successfully in the conditions of dark environment and rainy days or foggy days. Moreover, our system is able to run on the computer without any GPU to recognize the traffic sign efficiently and achieve the accuracy to 100%.