| 摘要: | 近年來為了減少交通事故而發展的車輛輔助安全駕駛議題越來越受重視。對駕駛人而言，在車的兩側都有一些無法觀察到的盲點視線範圍。我們可以在左右兩側後照鏡的下方架設相機拍攝側後方的影像，利用電腦視覺方法偵測是否有可能造成威脅的來車，以輔助駕駛人變換車道。 我們的盲點範圍視覺偵測系統包含：近側車道線偵測、虛實線判斷、遠側車道線位置估計、側方車輛偵測、側方車輛距離估計、側方車輛追蹤及車輛相對運動關係分析等模組。 以定義好的車道線模式，尋找最符合該模式的直線作為近側車道線。由已知的道路寬度和相機參數以反投影法估計出遠側車道線的位置，並藉由近側及遠側車道線所劃分出的範圍，作為偵測車輛的搜尋範圍。我們以車底陰影和車輛左右垂直邊作為偵測車體區塊的特徵，並藉由一些條件來驗證偵測區塊。我們會追蹤各個車體區塊以獲得同一台車在連續影像中的相對位置，並記錄該車輛的運動向量。如果車輛是朝己車逐漸逼近，則系統會對駕駛人提出警示，以避免切換車道造成碰撞。 我們以各種不同天候狀況和不同道路的影像測試偵測效能。由實驗結果顯示，在良好天候狀況下的側邊車輛偵測率約為92%，但在惡劣天候狀況下的偵測率並不理想。因此，研究新的偵測方法或是融合其他的感測器來幫助惡劣天候下的偵測是未來主要的研究目標。 Developing a real-time automotive driver assistant system for safety has emerged wide attention in recent years. When driving on the road, the fields of view beside the host vehicle for drivers are limited. Therefore, we utilize cameras mounted under side-view mirrors of a vehicle to monitor the circumstance in the blind-spot areas for drivers to avoid possible collision when changing lane. The proposed visual blind-spot detection system includes near lane mark detection, classification of solid/dashed lane mark, far lane mark estimation, side vehicle detection, distance estimation of side vehicles, vehicle tracking, and object-based motion analysis. In the proposed system, the lane mark at the near side of the host vehicle is detected by searching the optimal parameters of a defined lane model on the images, and the lane mark at the far side is estimated from the relative position of near lane mark by inverse perspective transform. Thus the proposed system is able to extract the region of the adjacent lane and detect the approaching vehicles. Side vehicles are detected by underneath shadow and left/right borders, and verified by the ratio of vehicle width and road width, symmetry, and gray-level variance of the vehicle region. We track the detected vehicles in consecutive images to acquire their relative positions between frames and compute their motion vectors. We analyze the motion vectors to judge if the vehicle is approaching the host vehicle, and the system will warn the driver if there is a vehicle approaching during the driver changes lane. In the experiments, we evaluate the proposed system in different weather conditions, such as cloudy day, sunny day, dusky day, and rainy day, and in different driving environments, such as highway, expressway, and urban roads. The average detection rate of vehicles in sunny day and cloudy day is about 92%, while the detection rate in rainy day is about 75%. The performance of the vehicle detection is not robust enough in bad weather condition, so finding other vehicle detecting method or fusing different sensor data is our future work. |
