以影像辨識技術為基礎之智慧型車輛偵測系統

近年犯案類型多元，傳統偵辦方式已無法滿足現況，因此內政部積極導入科技辦案，不僅加快破案速度且更加清楚釐清案發經過，經統計資料顯示2010年以影像技術破案比率佔5.34%，至2016年成長至17.81%，有鑒於汽車與監視器數量逐年提高，調閱影像搜查過程耗時且事倍功半，故本研究以EMGU CV影像處理技術為基礎，結合YOLO V2深度學習技術設計一套「以影像辨識技術為基礎之智慧型車輛偵測系統」，此系統主要包含三個模組：(1)車輛影像辨識模組－針對影像檔內容，將辨識車輛之參數儲存至資料庫；(2)車輛顏色搜尋模組－依據設定條件之RGB參數，於車輛圖檔資料集中進行比對，並顯示符合條件之項目；(3)車型搜尋模組－依據辨識結果資料庫中紀錄之長寬進行篩選，並將結果以清單方式呈現。本研究實驗結果車輛辨識準確率97.16%，顏色搜尋結果在資料前20筆中準確率為96%，車型搜尋結果在前20筆中準確率為87.5%，實驗結果顯示本系統具實務應用。

關鍵字

影像辨識； YOLO ；車輛偵測； EMGU CV

並列摘要

In recent years, the classifications of crimes are multiple, and the traditional ways to hand criminal cases cannot be satisfied with the circumstance at present. Consequently, the department of interior has actively led to high-technological to take charge of cases, which is not only speeds up the rate of solved criminal cases, but also specifies the procedure. Based on the information shows that, the solved criminal-case ration of video image techniques was 5.34% in 2010, and it grew to 17.81% in 2016. Because the quantity of vehicles and monitors increases year by year, the courses of having access to all surveillance cameras make double results with half effort. This study researched on the EMGU CV based on the image processing techniques to combine YOLO V2. By deep learning to design a suit is that “An intelligent car detection system based on image recognition technologies”, which is including three modules. First of all, a vehicle-image recognition module, according to the content of image files to store recognized vehicle parameters to data base. Second, a vehicle-color-searching module, on the basis of setting conditions in RGB to collect vehicle graphic information to image matching and show matching conditions. Finally, a vehicle-models-searching module, according to the outcome in data base recoding length and width to filter and show the consequences by a list. This study showed that the recognition accuracy of outcome reaches 97.16%, the color-searching accuracy of outcome reaches 96% on the top of 20 rows, the vehicles-models-searching accuracy of outcome reaches 87.5% on the top of 20 rows. The experimental result showed this system is practical application.