[[alternative]]多樣性的道路線及前車視覺偵測技術

[[abstract]]隨著經濟的成長，人民所得提高，生活品質提昇，機動車輛數量亦隨之快速成長。車輛增多，再加上開車時的各種環境、道路、及人為等因素，而使得交通事故的件數及其所造成的死傷人數也相對增加。因此若能有一些協助駕駛人偵測道路狀況、行車狀況、及駕駛狀況等機制，提供各種道路相關資訊給駕駛人，並提醒駕駛人，對社會大眾而言，勢必多了一層安全保障。過去我們已發展了一個單眼電腦視覺偵測系統，用於偵測及追蹤前車及道路線，分析前車與已車的行車狀況，確保駕駛人的安全。在本研究中我們改進了部份技術並增加了新功能，使得本系統擴充成有十二項偵測及分析功能：車道線偵測、多車道估計、虛實道路線分類、黃白道路線分類、單雙道路線分類、己車方向估計、己車左右位置估計、偏離車道警示、前車偵測、前車距離估計、煞車燈偵測、及方向燈偵測。其中雨天的前車偵測與追蹤、黃白道路線分類、單雙道路線分類、解決一般道路路口無車道線的問題、和解決夜間駕駛路燈發散現象問題是本論文所提供的。在雨天前車的偵測中，我們以車子底部和輪胎深顏色的特性，設計樣板並藉由樣板匹配來偵測前車的可能位置；去除樣板重疊加快系統執行效率和增加執行果。之後，再利用車寬和車道寬的比例、對稱性、及區塊色彩標準差來驗證是否真的是車輛。在道路線偵測方面，根據所得到的道路線兩側位置，取出道路線中線的顏色資訊，藉此區分黃白及單雙的道路線，再利用影像連續性的觀念，記錄每張影像中道路線的角度及截距，並且計算其變化程度並修正，藉此判斷是否為相機震動或是遇到路口。而利用道路線寬的觀念，來克服夜間駕駛時，因前車車燈或是路燈產生的光源發散現象。我們以多種影像：晴天、陰天、多雲、雨天、傍晚、夜間在一般道路及高速公路等影像，來測試我們的偵測效能。平均前車偵測率達94.5 %，在Pentium® 3.0 GHz 的電腦上，12 項功能全部執行的時間為0.045 秒。從實驗結果顯示，我們的系統可以在不同的天候及駕駛狀況下，即時且快速的偵測前車及道路線。Developing automotive driver assistance systems to alert drivers for possible collision with other vehicles or obstacles has attracted much attention lately. To achieve such a system, we use a camera mounted on a vehicle to capture road scenes for detecting the driving situation and environment. We had proposed a monocular vision system to assist the road driving for safety. In this study, we improve the proposed system and add new functions on lane and preceding vehicle detection such that thesystem consists of twelve functions: lane detection, multiple lane estimation, solid/dashed lane-mark classification, yellow/white lane-mark classification, single/double lane-mark classification, vehicle direction estimation, vehicle lateral offset estimation,lane departure warning, preceding vehicles detection, distance estimation for preceding vehicles, brake light detection, and turn signal detection. In this study, we detected the lane marks by searching the best-fitted parameters of a defined lane model on images, and we verify the results by the width of lane marks and the difference of the slopes and the intercepts in consecutive images. Preceding vehicle detection consists of two independent modules. In the first module, we utilize cast shadow and left/right borders of vehicles to detect preceding vehicles. In the second module, we use a template matching to extract vehicle candidates in the rainy days based on several templates generated from the images of tires and the bottom of vehicles. Following both modules, we use ratio of road and vehicle widths, symmetry, and gray-level variance to verify the candidate vehicles. In the experiments, the proposed system was evaluated in several different weather conditions: sunny day, misty day, dusty day, cloudy day, rainy day, and night. The average vehicle detected rate is 94.5 % and average processing time for executing all twelve functions is about 0.045 second run on a general PC with Intel® Pentium® 3.0 GHz CPU. From the experimental results, we find that the proposed approaches can stably detect and track the lanes and preceding vehicles in real time