使用路側低角度攝影機的車流量偵測系統

Abstract

近年來，基於影像式的車輛偵測技術在智慧型運輸系統中受到廣泛的重視與研究。然而，前景破碎與車輛遮蔽問題依舊是造成嚴重的錯誤與誤判的主要原因，如何改善前景破碎與車輛遮蔽問題所造成的影響，將是本篇論文努力的方向。在本論文中，我們提出一個嶄新且獨具創意的系統，有別於傳統利用影像畫面做處理，我們則是將影像轉換成俯視影像，在一斷面內配適當大小的感應區塊，將前景物特徵加以收集並與時間整合成一維前景序列，並以此一維前景序列代表影像畫面中的前景物資訊進行後續的處理流程，此方法將有助我們改善物體和背景相似所造成的前景破碎問題，俯視影像搭配完整的前景物資訊亦有助於我們制訂條件來判斷遮蔽問題的發生進而加以處理。我們提出的方法適用於道路側向的監控環境，高速行進的車輛依舊能有效分辨，也因此大幅提高車輛計數的準確率，在我們的系統中，白天的偵測率可以達到97.4%，夜間的偵測率依舊能維持97.07%的高水準。 本研究的目標在於改善影響車流輛偵測的兩大要素，第一，因環境因素導致物體資訊減少造成誤判或遺漏。第二，車輛未進入畫面前就發生遮蔽問題而離開畫面依舊存在遮蔽問題的狀況。由此系統所得到的結果可以在車輛偵測的技術中，針對以背景為基礎的車輛之偵測系統所遭遇到的問題提供另類的思考模式。實驗的結果證實我們提出的系統可以處理一定程度的遮蔽狀況，並偵測出影像中的車輛，我們所實作的系統也能為後續處理，提供有用的資訊。

Vehicle detection techniques in visual-based Intelligent Transportation System (ITS) have been studied for years. However, fragmented foreground and vehicle occlusion problem still cause serious errors and mistakes. How to reduce the influence of fragmented foreground and vehicle occlusion is our essential target. This study proposes a novel method and different from conventional processing on image plane. We transform image into bird-view image and collect foreground information from a cross-section by suitable sensing blocks. The data we obtain will transform into one-dimension feature sequence and represent the objects in future processes. This method will prevent fragmented foreground. Based on bird-view image and complete foreground we can design general conditions to detect occlusion cases and handle. Our system is capable of vehicle detection with roadside low-angle camera and with high speed moving objects. This method is sufficiently robust to detect vehicles at a high success rate of 97.4% in day and 97.07% in night. Our research is to reform fragmented foreground and handle vehicles that move in parallel occluding each other from the beginning to the end of the traffic image as a single region. And this study can provide a different concept from traditional background based method. Experimental results show the opportunity of tracking and counting systems and advanced applications.