| 摘要: | 由於經濟蓬勃發展，國內自小客車持有率逐年提高，駕駛者違規所衍生的交通問題日益嚴重。目前國內執法工作大多由警員人力支援，受限於人力資源，常無法達到常態性以及公平性…等原則，因此儘管加強違規的取締，成果卻不彰。透過影像處理技術來辨識車種以及取得交通參數，近年來於國內外已逐漸發展成熟。透過整合之智慧型交通執法系統，可以讓交通稽查取締工作發揮更大的功能，提升交通績效。
本研究利用影像處理技術結合自行發展之道路環境辨識演算法建立智慧型執法系統之平台。研究計畫分為三個部份：背景建立與更新、道路環境辨識以及違規行為偵測與辨識，並選定「未依規定變換車道」以及「未依規定行駛車道」違規行為建立違規行為判定子系統，做為智慧型執法系統平台之先驅研究。
本研究利用影像中位數法建立背景、結合遞迴式背景建構模式與非遞迴式背景建構模式進行背景更新；並且利用特徵判斷與樣板比對辨識路面標線，進而分割道路環境；最後利用車輛偵測、追蹤等…影像處理技術結合道路環境辨識結果進行違規行為的偵測與辨識。
經過實際車流影像驗證，發現利用影像中位數法所建立之道路環境背景於不同的時間點、不同的路口或路段，皆能建構出乾淨的背景。在道路環境辨識方面，利用特徵以及樣板比對進行判斷之路面標線辨識率分別為100%以及89.9%。車流量計算方面，路口內大型車、小型車以及機車的辨識率分別為84.2%、97.1%以及100%，；路段內大型車、小型車以及機車的辨識率分別為89.4%、98.3%以及97.7%。違規行為辨識方面，未依規定變換車道辨識率為100%、未依規定行駛車道辨識率為89.1%。上述之結果，雖然未達百分之百的辨識率，然已成功地建立可用的平台。
Because the highly economic developing for Taiwan within the past several years, the car-hold-rate is also increasing rapidly year by year. Therefore, the traffic problems which are from driver traffic violation are more serous than that before. Nowadays, there are many non-normal and unfair cases for the traffic management cause by the limited police resources. Although the police agency enhances to suppress illegal uses of traffic, the performance is still not good. The intelligent enforcement system platform is used to improve the traffic violation detection performance.
In this study, the image processing techniques and the road environment recognition algorithm are used in the intelligent enforcement system platform. There are three parts: 1) the background reconstruction and its updating. 2) The road environment recognition. 3) Traffic violation detection. Here the “Change-Lane at Will” and the “violation of Right-Lane Driving” are selected for the traffic violation detection.
In our study, the background is constructed using temporally median filter and the combining it with recursive and non-recursive background updating algorithms to update our background image. Next, the extracted features and template matching algorithm are used to obtain the lane edge trace. Hence, the road environment can be segmented via these known lane edge trace. Finally, the moving car detection and its tracking algorithms are also used to recognize the traffic violation.
The simulation results show that the temporally median filter can construct a clear background image even in the different environments. In the road environment recognition, the detection rates of the proposed feature extraction and template matching methods are about 100% and 89.9%, respectively. In the traffic-flow counting, the detection rates of big-car, small-car and motorcycle are 84.2%, 97.1% and 100%, respectively. In the traffic violation detection, the detection rate of “Change-Lane at Will” is 100%, and the detection rate of “violation of Right-Lane Driving” is 89.1%. Althourh the detection rates were not achieve 100%, but the feasible platform was established successfully. |
