隨著近年來汽機車的普及化，而汽機車的數量逐年的提升，然而因為汽機車的數量提升導致交通事故也逐年提升。因此先進駕駛輔助系統(Advanced Driver Assistance Systems，ADAS)被廣泛的使用在汽車上，希望藉由駕駛輔助系統來降低交通事故。 本篇論文關注於駕駛輔助系統中，基於車門後方影像進行事件警示的預防碰撞系統。在學術領域研究方面，大多使用紅外線感測器偵測後方車輛，根據後方車輛與自身車門的距離決定是否警示。然而，這種方法的缺點是當感測器發出警訊時，但此時後方的來車是往遠離車門的方向前進，因而造成誤判。 本篇論文提出一個較彈性的系統架構，自動偵測車門位置，並定義出感興趣區域，偵測區域中後方來車。希望手機或相機架好之後便能夠自動偵測車門後方危險來車。 本論文所提出的系統不依賴感測器偵測方法，而是使用分析相機或手機影像的稠密光流(Dense Optical Flow)作為車門後方事件的特徵。我們以移動物的移動軌跡進行分群，針對每一群軌跡找出特徵向量，並利用自適應增強演算法(Adaboost)和支持向量機(Support Vector Machine)分類器判斷危險事件是否發生。 在實驗中，我們展示本論文提出的系統對於車門後方危險事件偵測有良好的可靠度，並且在影片中危險的事件都能準確的偵測到，達到一台手機或攝影機能夠偵測車門後方來車危險事件之目的，此系統在個人電腦上達到每秒鐘29真幀的實時運作。 ;With the popularization of cars and motorcycles in recent years, the number of cars and motorcycles increases year by year. However, the increasing number of cars and motorcycles results in car accidents year after year. Hence, Advanced Driver Assistance Systems is widely used on cars. Wish to reduce the car accidents via the Driver Assistance Systems.
This dissertation focus on Driver Assistance Systems, which is based on analyze source images and detect vehicles in the rear. In academic research, it mostly uses infrared sensors to detect the tailing cars and decides whether to alarm according to the distance between the tailing cars and our own car doors. However, the defect of this method is that it may misjudge because while the sensor is alarming, the approaching car from the rear is moving away from the car door.
This dissertation proposes a rather more flexible system structure. It can sense the position of car door automatically and define a specific area to sense the approaching cars in it. Wish that once a cellphone or a camera is erected, the system can sense dangerous approaching cars from the back of car doors automatically.
The system in this dissertation does not depend on sensor sensing method but analyze the Dense Optical Flow of cameras or cellphone videos as features of incidents at the rear of car doors. We group the tracks of moving objects and find out the eigenvector of each group with the tracks and utilizing Adaboost and Support Vector Machine to determine whether dangerous events will happen.
In experiment, we display that the system, which is proposed in this dissertation has great reliability on detecting dangerous things in the back of car doors. Furthermore, in the video, dangerous events can be detected precisely, achieving the goal that a cellphone or a camera detects the dangerous approaching cars from the back of car doors. From practical operation, this system achieves 29 frames per second on personal computer.
