多視角車牌影像融合技術之評估

Abstract

近年來，視訊監控的發展將會朝向以多鏡頭、多視角的方式作攝影，以提供使用者更豐富、更精細的視訊監控效果。而新一代視訊監控的技術關鍵就在於多視角影像融合技術，若能有效融合多視角影像間的資訊，便能大幅強化視訊監控的功能，進而提昇視訊監控的價值。如何善用多重影像來源之間的額外資訊，提供更精確的分析與辨識，以及高品質呈現，是重要的研究課題。本論文的目的，將以多視角車牌影像(Multi-view License Plate Images)為主要輸入與處理對象，探討基於平面轉換的估算，融合多視角影像之間的資訊，以得到更高畫質的車牌影像，並分析造成影像融合誤差的因素。實驗結果顯示，基於平面轉換的車牌影像融合技術確實可以有效提升車牌字元的影像品質，並有利於後續車牌辨識的應用。

It is demanding to acquire high quality images for various applications in visual surveillance. To achieve the goal, a new trend is to integrate information from multi-view images captured by multiple cameras. With proper image fusion, not only the visual qualities can be enhanced but the performance of subsequent image recognition can be improved. Thus, developing effective techniques for image registration and information fusion of multi-view images is very important. In this thesis, we present a technique of fusing multi-view license plate images based on homography transformation, and investigate the factors which result in fusion errors. Experimental results show that by proper selection among multi-view images, using both simulated and real image data as inputs, reasonable fusion of license plate information can be obtained from images rectified using homography.