| 摘要: | 隨著經濟快速成長，人民所得提高，生活品質提昇，人口增加，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，而使得交通事故之件數及其所造成之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。本研究的目的即是結合電腦視覺、影像處理、與圖形識別等技術，考慮各國環境之需求，發展與世界同步之視覺偵測技術，給予駕駛者提醒、警示、及預防等機制，降低駕駛者工作負荷，減低人為失誤，進而預防交通事故發生，提升行車安全性。從 2001 年起，我們就一直在從事先進安全車輛之視覺偵測技術的研究。藉由多樣性的電腦視覺偵測技術，偵測駕駛環境及個人狀況，提供警示功能，以提高行車安全。本研究即建立在過去的基礎及成果上，繼續針對各種可能的交通安全性及駕駛方便性，發展更新的或更穩定的電腦視覺輔助道路安全駕駛技術；讓技術得以實現出來，創造利潤，造福人群。我們過去的研究重點在於單張影像的處理，在本研究中，我們將研究重點擺在連續影像與立體視覺的研究上。我們所擬定的研究內容有: 1. 單眼連續影像的動態偵測：i.側邊盲點區域的多重解析度光流 (optical flow) 偵測，ii.慢速三合一 (影像式停車導引、自我移動的障礙物距離估計、及移動視覺的立體障礙物判定) 的特徵比對動態偵測，iii.前車停止與啟動的調適性光流偵測，iv.俯瞰監視的粒子濾波器 (particle filter) 動態偵測等四項研究議題。 2. 連續影像的穩定分析：在連續影像的視覺偵測中，一定存在偵測不穩定的問題；例如，i.前車偵測與測距的結果，ii.側邊盲點車輛的偵測結果，iii.倒車導引線條的穩定度，iv.雙眼立體視覺行人偵測的結果等四項連續影像的偵測常會有環境、亮度、及機械等因素的干擾而無法維持連續一致的偵測及距離估計結果，而給使用者系統不穩定及常出錯的印象。在本研究中，我們將以擴充式卡曼濾波器 (extended Kalman filter) 為主的連續動態分析，關聯前後影像的偵測內容讓連續偵測更可靠更穩定；並且加入學習的機制。 3. 雙眼立體視覺偵測：在視覺偵測中，常有因為物件過小或變形而偵測不穩定的問題。為了這些因素，我們擬議採用雙眼立體視覺偵測行人及單機車。立體視覺技術有許多種，但適合戶外長距離應用的只有雙眼立體視覺。雙眼視覺有繁複、慢速的問題。在本研究中，我們將發展平行處理晶片(GPU) 配合DSP 的立體視覺演算法。另外我們亦規劃三項可以與前述技術配合的後續技術：ii.以擴充式卡曼濾波器偵測駕駛人的臉部三維移動與轉動參數，iii.以擴充式卡曼濾波器平穩連續夜間車燈偵測結果，iv.以光流提昇乘客上下車偵測的正確率。上述三年研究除了電腦視覺技術發展外，還都會使用嵌入式系統及參考ISO 標準規格等議題。上述相關於安全偵測的研究重點在於準確性與穩定性。視覺安全偵測很容易受到天候及環境因素的影響；所以技術研發的關鍵在於方法是否不受天候狀況 (晴、陰、雨、霧、暗) 的影響；是否在各式道路環境 (高速、快速、平面道路、巷弄、鄉間小道) 都可維持良好的偵測效果。考慮各種天候及環境因素對於視覺偵測的影響是我們技術發展上的特色之一。計畫主持人已有二十多年電腦視覺的研究經歷，且已有近十年電腦視覺技術應用在車輛安全駕駛上的經驗；更從2008 年起合聘到工研院機械所智慧車輛組協助先進安全車輛之視覺偵測技術的發展，並且獨自與共同獲得或申請多個電腦視覺輔助道路安全駕駛之中華民國、美國、及中國發明專利，其部份成果具備高度實用性，已被數個單位及公司選用於發展產品；因此我們有信心及能力完成本計畫的執行。In these few decades, the vehicle number is rapidly increasing because people’s incomes are increasing. In addition to the vehicle number, more factors of road situation, driving environment, and human attention result in a large amount of traffic accidents and casualties. If there is a mechanism to help the driver to detect the road situation and driving environment, and then to provide some useful information to the driver in these situations, the danger is therefore avoided. In this research project, we are just combining the techniques of computer vision, image processing, and pattern recognition, and matching the domestic environment to develop state-of-the-art techniques of vision detection to reduce the load of drivers, to reduce the traffic accidents, and then improve the driving safety. From 2001, we have studied the vision detection techniques for the advanced safety vehicle. Based on the versatile computer vision techniques, we proposed methods to detect the situations of driving environment and personal conditions and provide warning to improve the safety of driving. Based on our previous basis and the results, this research project aims various driving conditions to develop new-function and more practical techniques to aid driving. In the previous study, we focus on the development of single-image processing techniques. In this study, we will pay all effort on the processing techniques of sequence and stereo vision images. We propose three studying topics: 1. Motion detection for sequence images: i. multiresolution optical flow estimation for blind spot detection, ii. three low-speed feature-matching-based motion detection, iii. adaptive optical flow estimation for stop-and-go application, iv. motion detection for surrounding top-view detection based on particle filter. 2. Stability analysis for contiguous image sequence: In the vision detection of image squences, there always unstable problem of detection results; such as, the results of i. detection and distance estimation of preceding vehicles, ii. blind-spot detection, iii. guiding lines for parking guiding, iv. pedestrian detection based on the binocular stsreo vision, are always unstable due to the interference of environment, illumination, and other factors. In this study, we will use the extended Kalman filter (EKF) to enhance the motion detection results. 3. Binocular stereo vision detection. Sometime, targets are hard to be detected due to their sizes and appearances. Thus we will use binocular stereo vision method to detect pedestrian and bikes. Binocular stereo vision has the problems of complexity and time consuming. In this study, we will develop the parallel algorithm based on the graphic processing unit (GPU). We also create other three topics in this year, which are ii.3D location and orientation estimation for driver face based on the EKF, iii. Improve the vehicle detection results in night based on EKF, iv. increase the passenger detection rate on a taxi using optical flow information. The studying items in each research topic will be sequentially completed in the following three years. The proposed system seemingly includes too many items; however, most topics are extended form our previous studying results. The key problems of the safety detection techniques are accuracy and stability. The methods concerning the safety vehicles are easily influenced by climate conditions and environment factors; thus the key points of useful methods are whether the proposed methods are not influenced by various climate conditions (sunny, cloudy, rainy, foggy, and night days) ? whether the proposed methods may keep well performance in variant road conditions (freeway, highway, general road, and country roads). Considering the influence factors of climate conditions and environment situations to the vision detection techniques is one characteristic of our development of the computer vision techniques for advanced safety vehicles. The principal investigator of this project is an original researcher on computer vision, he has studied computer vision techniques more than twenty years; moreover, he has the application experience of computer vision aided road vehicle driving for safety more than nine years. From 2008, he has also been a faculty in the Intelligient Mobility Technology Division, Mechanical and Systems Resaerch Lab., ITRI to help the development of the vision detection techniques for vehicles. He has also gotten and applied the US, Taiwan, and China patents in these few years. Partial techniques of the system are practiced and have been employed by several companies; thus we have ability to complete the execution of the research project. 研究期間：10008 ~ 10107 |
