交通監控影片中日夜車流壅塞之分析及評估

Abstract

近年來，各國致力於智慧型運輸系統的開發，以期能透過即時的交通訊息傳輸與整合來提升交通運輸品質。對於一般大眾，若能從中獲得即時的交通壅塞資訊則是更加的實用。目前，由攝影機所構成的交通監控系統已成為交通訊息偵測的主流，然而大部分的研究僅限於監控影片中交通事件的自動分析，例如交通事故與違規事件，但這無法幫助我們得知當下車流的壅塞情形。所以，處理監控影片以提供民眾最即時的交通壅塞資訊是迫切需要的。在本論文中，我們提出了一個適用於白天與晚上交通監控影片的壅塞程度評估系統能將壅塞程度分為五個等級。 想要從影片中評估交通壅塞程度，視訊處理的技巧與相關的知識是不可或缺的。對於白天的影片，我們利用背景相減法來找出道路上的車子，而在晚上的影片中，則是透過車頭燈的偵測來找出車子的位置。當車子擷取出來之後，我們使用虛擬的偵測器來蒐集交通訊息，藉此即時地評估所偵測到的車流壅塞程度。除此之外，我們也針對道路偵測、車流方向判斷與車道偵測等問題提出解決方法，以提升整個系統的即時性與完整性。最後，我們利用高速公路上的交通監控影片來驗證我們所提出系統的性能，並獲得了令人滿意的結果。

In recent years, intelligent transportation system is developed to promote the quality of the traffic transportation. In general, concerns of the traffic control center are traffic management, vehicle control, and traffic safety. However, they are not the issues that people concern most. Instead, the situation of traffic congestion is much more useful for the public. In addition, traffic surveillance systems have been widely used for monitoring the roadways. There have been many researches on video analysis of traffic activities such as traffic accidents and violations, but these researches still cannot help people get to know the traffic congestion situation. Therefore, we intend to develop techniques to process traffic surveillance videos for providing people with instant traffic congestion information. In this thesis, a traffic congestion classification framework is proposed for identifying the traffic congestion levels in daytime and nighttime surveillance videos. The degrees of traffic congestion are classified into five levels: jam, heavy, medium, mild and low. In order to analyze the traffic congestion levels from videos, image processing techniques and the knowledge of classification are indispensable. In the proposed framework, moving vehicles are extracted by background subtraction during the day and by headlight detection at night. Afterward, virtual detectors and virtual detection line are utilized to evaluate and classify the traffic congestion levels in daytime and nighttime surveillance videos, respectively. Moreover, methods of bidirectional roadway detection and lane detection are proposed to extract the consistent features of roadway for the requirements of real-time response and robustness of the frameworks. In the experiments, we use real freeway surveillance videos captured at day and night to demonstrate the performances on accuracy and computation. Satisfactory experimental results validate the effectiveness of the proposed framework.