在本篇論文中,我們提出了一個系統化的夜間車輛辨識與預測系統。車輛偵測在智慧型交通系統這個領域中,一直扮演著很重要的角色,而且在夜間行車環境視線較差,駕駛人更加需要輔助系統提供行車安全資訊。因此,在夜間環境下確實的找出車輛的位置,成為了很多研究中的重要議題。最近幾年,許多研究致力於對車輛行為做分析,以預防交通事故的發生,並輔助駕駛行車上的安全。然而,在夜間環境下透過行車記錄器進行車輛偵測,比起在日間光線充足的環境下要來得更具挑戰性。在本論文中,我們使用了CenSurE去檢測可靠的可能車燈,並加以配對,為了降低因大量錯誤配對而可能引發的誤偵機率,我們使用了一套透過車燈分布模擬的過濾方法,可以有效濾除因其他燈光或鄰近車輛間的錯誤配對。接著我們提取車燈部份的HOG特徵透過SVM分類器分類出車輛跟非車輛。此外,我們也提出了一個利用車輛假設來預測車燈位置的機制,重新找回因缺少車燈或分類器漏偵而無法偵測的車輛。只要該車輛的任一車燈能被偵測到,我們都可以持續偵測出該車輛來 。實驗結果顯示我們所提出的方法在夜間行車環境下能夠很好的去辨識出車輛的位置。
In this study, a systematic method for automatically detecting vehicles in nighttime is proposed. In nowadays, vehicle detection plays an important role in the research domain of Intelligent Transportation System (ITS). Due to the poor visibility in nighttime, drivers need more on-road information, e.g. the distance to the car in front, to improve driving safety. Many researchers make a great effort in this field. However, it is still a challenge task to detect on-road vehicles by camera since tail lights are the only visible features of a vehicle. In this work, we use CenSurE filter to detect reliable light sources. To identify a car by its tail lights, we pair the detected lights and filter out improper pairs by the distance between two lights. Then, each light pair determines a Region of Interest (ROI). We extract Histogram of Oriented Gradient (HOG) feature from the ROI and use Support Vector Machine (SVM) to classify it into vehicles and non-vehicles. We also develop a vehicle hypothesis to predict the position of the vehicles whose light is miss detected by CenSurE filter. Our experimental result demonstrates that the proposed method is effective and efficient in practice.