具有環境適應性之即時遠端車輛偵測系統於複雜交通情況之研究

Abstract

一般常見的車輛偵測器通常使用收斂背景的方式來偵測車輛，在得到了影像中的背景影像之後，再用整張影像減去背景影像後得到的前景影像，來完成移動物體的偵測。對於影像中光線變化較為急促導致偵測出現錯誤的情況，則通常是使用背景更新來解決；但若是在車輛壅塞時，由於路面上的車輛移動皆極為緩慢，故此類系統在收斂背景時，容易將緩慢移動甚或是靜止的車輛收斂為背景，導致原影像減去背景影像時，所得到的前景影像出現錯誤，而造成車輛偵測的失敗。此篇論文主要的貢獻在於提出了一種能夠動態適應各種光線環境，且在車輛雍塞時能增加車輛偵測準確率的影像式車輛偵測器。且在提出的系統中，也獨特的利用了畫面中顏色的強度差距將不屬於路面的移動物體擷取出後再利用追蹤的方式提昇系統的偵測準確性。除此之外，我們也將此種車輛偵測器與影像壓縮及網路傳輸結合，使之能將CCD攝影機架設地所攝入的畫面即時的壓縮並經由網路傳輸到後端處理中心做車輛偵測的動作。此種應用在未來更可結合PDA或是行動電話，經由遠端行車資訊的獲得，以提供行車用路人各種的行車資訊。

Vehicle detectors utilize the background extraction methods to segment the moving objects. The background updating concept is applied to overcome the luminance variation which always results in the error object detection. These systems will be challenged to detect the vehicles in traffic jams or in different luminance conditions. Since vehicles will cover the road surface or move slowly in traffic jams, the background is difficult to be converged or updated. Once the traffic is released, the existed background is unsuitable for the follow-up moving object segmentation. In this thesis, an adaptive vehicle detection approach is proposed to improve the detection accuracy in traffic jams. And a histogram extension method is utilized to segment moving objects robustly with luminance adaptation. Furthermore, the vehicle detection based on merged boundary box rule method is a good solution to reconstruct the broken moving contour for enhancing the detection accuracy. And the moving object segmentation method based on the histogram extension process is presented. Besides that, the detection system detects the vehicles from the image sequence which captured from the CCD camera on the road side and transmitted via network real time after compressed.