偏振片標籤偵測與識別系統

Abstract

本論文利用偏振片及可偵測偏振光之特化相機，實作出長距離定位及辨識用途之標籤系統。在一般的影像辨識標籤系統中，被偵測標籤往往利用顏色、大小或其他不規則圖樣或花紋當作特徵以供偵測。然而此作法在距離增加後，由於影像解析度下降而無法辨識出來，辨識度將會迅速下降。為了克服上述困難，除了空間上的變化之外，此系統導入了偏振光的特性，利用環境中偏振光線較少的特性，減少環境光所帶來的雜訊，藉此提高標籤在圖像中的訊噪比，以供長距離時仍可迅速的被系統所偵測並解碼。 我們實作了一個可供長距離偵測及辨識系統。此系統在標籤部分採用多片帶有不同偏振角度的偏振片，以不同角度的組合做為識別用的調變；在讀取器部分，則採用了帶有特化後的感光元件的相機，該感光元件表面覆蓋多個不同角度的偏振片，使得相機得以拍攝出在不同偏振角度下所觀測的影像。藉由偏振相機所取得的偏振影像，可供計算出該影像中的偏振資訊，此資訊得以使系統過濾大部分非偏振的光線，使得感興趣區域僅剩數個較強程度的偏振光區域，藉此減少整體偵測及辨識的運算量，另使用粒子濾波器增加標籤追蹤的效率。 此系統目前在室外，讀取器與偏光標籤相距20公尺處，能以86% 的準確度辨識出16個不同身分，並以100% 的正確率辨識30公尺處辨識出4個不同的身分，其中標籤僅佔畫面中9*9個像素點。

Among modern object tracking techniques, in addition to utilizing machine learning to directly recognize the object’s shape, special pattern of markers or radio frequency identification (RFID) are often used to locate the object with attached markers. However, existing solutions have several drawbacks. Machine learning relies a large amount of learning data and massive computation resources. Traditional fiducial markers are limited by observation distance or angle, and need to occupy space with visually intrusive pattern. RFID cannot provide the fine resolution of spatial domain. In this paper, we turn away from traditional marker which consists of a number of black and white blocks in different sizes. Instead, we use polarizer to compose our markers, and use polarization camera to detect the marker and decide its identification number. Human eyes cannot distinguish light with different polarization properties, which implies this marker can be attached to many different surfaces which contain other visual information such as street name plate or traffic sign. However, unpolarization light can be filtered out with through polarization camera, which can separate polarized light from ambient light easily. This allows P-tag to be recognized even when it occupies just tens of pixels in the captured image. As the polarization camera can accurately determine the orientation angle of a particular polarizer, polarizers in different angles can be used to represent larger amount of information. With a marker consisting of a small number of polarizer “blocks”, we can represent a large number of identification numbers, e.g., 3 by 3 marker can represent approximately 50M ID numbers. Besides, P-tag is completely passive and battery-free. On the camera end, we use particle filter to detect and track markers. Particle filter can achieve the cycle of “prediction — observation”, which allows us to perform recognition based on the last observation result. Particle filter also provide resilience marker’s perspective distortion since particles just evaluate points which relative to it.