影像處理與電腦視覺技術應用於駕駛輔助系統之研究

Abstract

本論文主要探討應用於駕駛輔助系統之影像處理與電腦視覺技術，包括車道偵測、車輛偵測、前車距離估測、誤差估測及攝影機動態校正。電腦視覺為基礎(Vision-based)的駕駛輔助系統利用安裝在智慧車內的攝影機拍攝前方路況，透過車道與車輛偵測技術估測車道位置、前方車輛與智慧車的距離，這些資訊可以用來提高駕駛安全。本論文主要包含三個部份，第一部份簡介電腦視覺技術應用於駕駛輔助系統。第二部份為分析偵測道路所得的資訊及降低誤差的方法，第三部份則提出一些演算法，以應用於估測車輛距離及誤差、動態校正、以及車道與車輛偵測。 本論文提出一些新的方法來估測前車與智慧車的距離、快速估計前方物件的大小、距離與物件大小估測結果的誤差分析、及動態校正攝影機參數以降低誤差。首先，利用攝影機模型將世界座標的地平面座標轉換成影像座標，用以估計前方車輛與攝影機的相對位置，然後，利用一個新的估計方法估計前車的大小與投射的大小。這個方法利用前方車輛輪胎與地面的接觸點估計前車與攝影機的距離，並且利用前方車輛的其它頂點之投射位置，估計車輛的真實大小。因為前車投射的大小會隨著其與攝影機之距離而改變，論文中提出一個簡單且快速估計投射高度的方法，它能簡化繁複計算及降低計算時間，使得此設計能應用於即時處理系統。 距離估測的誤差分析結果顯示出當估測車輛與攝影機之間的距離時，攝影機的安裝參數也會影響到估測的結果，我們藉由誤差分析來找出最合適的攝影機參數以降低誤差。此外，因為車輛會因路況不平或載重不平衡而晃動或傾斜，安裝在車上的攝影機的外在參數也會隨著車輛的行進晃動而改變，因此，我們亦提出了動態校正的方法，以取得正確的攝影機參數，降低估測誤差。實驗結果顯示我們的方法可以準確的估測車輛大小及距離。 論文中也提出了一個快速估測投射的車道與車道線寬度的方法，用以預測車道的可能位置。另外，設計了一個車道線擷取狀態機Lane Marking Extraction (LME) Finite State Machine (FSM)，用以辨識影像中的車道線；並將cubic B-spline應用於曲線擬合以重建道路邊界。另外還發展了一個統計搜尋演算法用以決定在不同亮度條件下所設定的車道線擷取門檻值。此外，有時會有部份車道線被遮蔽而影響偵測，我們應用模糊演算法來判斷車道線可能被遮蔽的情形，進而利用已有的車道線資訊及估算的車道寬度補償被遮蔽的部份資訊。最後，為了加速偵測、減少偵測誤差及影像雜訊干擾，亦規劃一個ROI (Region of Interest)決定策略，它能提高偵測系統的穩健性並加快偵測速度。 另外，本論文發展了一個以模糊邏輯演算法為基礎的外形大小相似性演算法(Contour Size Similarity, CSS)。利用偵測和估測的影像中車輛大小的比較結果及模糊規則來辨識車輛。車輛偵測主要針對和智慧車相同車道的前方車輛。實驗結果顯示提出的方法可以有效的偵測車輛並估計其距離，而且當有車輛切入前方車道時，偵測的目標也會轉移到目前因切入而成為離智慧車最近的這一輛車。最後章節呈現了本篇論文的結論與未來的研究展望。

The dissertation aims to explore techniques of image processing and computer vision applicable to driving assistance system, including lane detection, vehicle detection, estimation of the distance to the preceding car, error estimation, and dynamic calibration of cameras. The vision-based driving assistance system films the front road scenes with a camera equipped on the intelligent vehicle, computes lane positions and the distance to the preceding car by the lane and vehicle detection and then adopts the obtained information to improve driving safety. The dissertation mainly includes three sections. The first section is a brief introduction of the application of computer vision techniques to the driving assistance system. The second section presents analyses of the information obtained from lane detection and approaches for reducing errors. The third section proposes some algorithms and their application to the range estimation, error estimation, dynamic calibration, and detection of lanes and vehicles. The dissertation presents several approaches to estimate the range between the preceding vehicle and the intelligent vehicle, to compute vehicle size and its projective size, and to dynamically calibrate cameras. First, a camera model is developed to transform coordinates from the ground plane onto the image plane to estimate the relative positions between the detected vehicle and the camera. Then, a new estimation method is proposed to estimate the actual and projective size of the preceding vehicle. This method can estimate the range between the preceding vehicle and the camera with the information of the contact points between vehicle tires and the ground and then estimate the actual size of the vehicle according to the positions of its vertexes in the image. Because the projective size of a vehicle varies with its distance to the camera, a simple and rapid method is presented to estimate the vehicle’s projective height, which allows a reduction of the computation time in the size estimation of the real-time systems. Errors caused by the application of different camera parameters are also estimated and analyzed in this study. The estimation results are used to determine suitable parameters during camera installation to reduce estimation errors. Finally, to guarantee robustness of the detection system, a new efficient approach of dynamic calibration is presented to obtain accurate camera parameters, even when they are changed by camera vibration arising from on-road driving. Experimental results demonstrate that our approaches can provide accurate and robust estimation of range and size of the target vehicles. In the dissertation, an approach for rapidly computing the projective lane width is presented to predict the projective lane positions and widths. Lane Marking Extraction (LME) Finite State Machine (FSM) is designed to extract points with features of lane markings in the image and a cubic B-spline is adopted to conduct curve fitting to reconstruct road geometry. A statistical search algorithm is also proposed to correctly and adaptively determine thresholds under various kinds of illumination. Furthermore, parameters of the camera in a moving car may change with vibration, so a dynamic calibration algorithm is applied to calibrate camera parameters and lane widths based on the information of lane projection. Besides, a fuzzy logic is used to discern the situation of occlusion. Finally, an ROI (Region of Interest) determination strategy is developed to narrow the search region and make the detection more robust with respect to the occlusion on the lane markings or complicated changes of curves and road boundaries. The developed fuzzy-based vehicle detection method, Contour Size Similarity (CSS), performs the comparison between the projective vehicle sizes and the estimated ones by fuzzy logic. The aim of vehicle detection is to detect the closest preceding car in the same lane with the intelligent vehicle. Results of the experiments demonstrate that the proposed approach is effective in vehicle detection. Furthermore, the approach can rapidly adjust to the changes of detection targets when another car cuts in the lane of the intelligent vehicle. Finally, a conclusion and future works are also presented.