以車道線偵測為基礎之駕駛人昏睡警示安全系統

Abstract

近年來，由於車輛數目的快速成長，造成交通事故問題日益嚴重。在台灣，每年有超過兩千五百人在車禍中喪生。而根據交通部的資料，過去的四年中，每年至少有八萬多件的車禍事故發生。這種情況下促成智慧型運輸系統(Intelligent transportation system, ITS)相關研究的發展越來越受到關注。而大部分車禍發生的原因主要由於駕駛人本身的分心、不注意車況、疲勞駕駛等不適當駕駛行為所引起。因此，為了能盡量避免駕駛者身處於此危險狀態，我們針對車輛後照鏡旁的側邊影像資訊，開發一套以智慧型視覺技術為基礎的車道偵測及偏移系統，以確保駕駛人行駛的安全性。 在車道線偵測部分，為了提高車輛側邊的視角範圍，我們將一支魚眼(Fish-Eye)攝影機架設在後照鏡下方，並利用車體資訊在連續影像中固定的特性，自動選取路面範圍，而不需要事先得知攝影機架設的相關資訊。為了可適用於全天候的光線變化條件，我們同時處理空間及時間軸上的影像資訊，使此系統在白天及夜間人眼可視車道範圍內，都能獲得清晰的車道邊界資訊。另外本論文提出一套分段直線搜尋模組來連結車道線的軌跡，以提升整個搜尋速度，並克服魚眼鏡頭失真的問題。 在車道偏移判斷的部分，本論文利用先前偵測的車道側向位置，及TLC(Time to Lane Crossing)的瞬時資訊，規劃車道偏移警示的觸發條件。另外建立一個可以即時更新的車道線位移穩定區間，來模擬駕駛人在直線道路行駛時和車道線保持習慣性距離的特性。最後我們和交通大學腦科學中心(Brain Research Center, NCTU)合作，取得其在虛擬實境動態模擬駕駛系統的環境下，針對駕駛人昏睡狀態預測的相關數據，套用在實際駕駛的影像內容中，使本系統除了估測車道偏移的外在因素外，也可針對駕駛人本身的精神狀況作更進一步的分析，以提升本系統的安全性及可靠性。 本論文發展的車道偵測輔助系統在1.83GHz的PC平台上平均可達超過15fps的執行結果。測試影像內容為在高速公路的實際駕駛環境，並在白天及夜間範圍內都維持穩定的偵測結果。

As the high growth of population of vehicles, the traffic accidents are becoming more and more serious in recent years. In Taiwan, more than two thousand and five hundred people are died in traffic accidents every year. For each of last four yours, the number of traffic accidents is at least eighty thousand according the statistics of the Ministry of Transportation and Communications (MOTC, R.O.C.). In this situation, a lot of researches about the intelligent transportation system (ITS) have been paid more and more attention to the researches of related fields. Most occurrence of the car accidents results from the distraction, inattention for the adjacent cars, and driving fatigue of the driver. As a result, to avoid the driver being in danger as much as possible, an intelligent vision-based system focused on image contents of lateral-view camera setting under the rear-view mirror on vehicle is developed about lane detection and lane departure warning in this study. In this thesis of lane detection, a fish-eye camera is located on the vicinity of the rear-view mirror to increase the range of lateral-view angle. Furthermore, we make use of the invariant of image for car body fixed in consecutive image sequences to extract the ROI (region of interest) containing the road surface without realizing the intrinsic and extrinsic parameters of camera in advance. To make this algorithm suitable for various light conditions all day, the information of image in spatial and temporal domain must be simultaneously processed so that the lane boundary keeps distinct whether people have seen in the day or night environment. On the other hand, a piece-wise line searching model proposed in this paper is to connect the trajectory of lane and to reduce the computation load and to overcome the fish-eye lens distortion. In the thesis of lane departure warning, the instantaneous information of the lateral position from the result of lane detection and the TLC (time to lane crossing) can be regarded as the warning triggers for the alarms of lane departure. Then, a stable-driving region with real-time update mechanism is constructed to simulate the straight-road driving habit of different drivers which get used to keep approximately the same distance between the vehicle and lane markers. Eventually, by cooperated with the BRC (Brain Research Center, NCTU), we utilize the statistics about drowsiness estimation of the drivers in Virtual-Reality (VR) dynamic driving simulator to implement in the video contents for realistic driving. Therefore, this mechanism can be not only estimated the external factors such as departure of lane boundary but the internal ones such as the conscious analysis of the driver with higher reliability and safety. The lane detection and departure warning system proposed in this paper has been successfully evaluated on the PC platform of 1.83-GHz CPU with the average frame-rate is up to 15fps. Moreover, this algorithm can be maintained stable results whether in the day or night environment of the realistic driving on highway.