使用偏振光之物件偵測及辨識系統

Abstract

本論文利用光的偏振性質及攝影機通訊，實作出一長距離物件偵測及辨識系統。在常見的物件偵測及辨識系統中，被偵測的物件通常都會藉由空間上的顏色變化，讓外觀上有特殊的圖案或花紋，藉此來增加物件偵測及辨識的準確度。然而此做法在距離增加時，由於物件在相機中的解析度降低，物體的辨識率會迅速下降。為了克服上述困難，我們提出了一個不同的解決辦法，除了空間上的圖案變化，也利用了時間上的強度變化，也就是光的調變。而為了增加偵測即辨識的距離，我們利用了能夠在長距離也能夠被接收，且強度變化不會被人眼所察覺的偏振光，使相機能夠在遠距離仍然能夠準確地進行偵測。 我們實作出了一個長距離偵測及辨識模板。此模板在偏光標記部分，使用了反光貼片、偏光片及液晶單元；在讀取器部分，則使用了電控偏光單元及相機。在偏光標記部分，系統會調變液晶單元，使其傳遞出帶有特定偏振訊號的偏振光，而此偏振訊號就代表著該偏光標記的身份。讀取器的電控偏光單元則是被用來分析環境中每個像素的偏光資訊。藉由偏光資訊，我們可以過濾環境中的雜訊，使得感興趣區域，縮小至數個只發出較強程度偏振光的區域，並藉此來減少整體偵測及辨識的計算量，增加整體的系統效能。透過此種方法，我們的系統能夠在使用幀速率100的相機時，達到即時偵測及辨識的功能。本系統目前在室外，讀取器與偏光標具距離30公尺處，能以86%的準確度辨識出16個不同身分，並以100%的準確度辨識出8個不同身分。而此時該偏光標記在影像中只佔了6乘6的解析度。

This thesis presents PoliTag, a long-range object detection and identification system using polarized light and camera communications (CamCom), which modulate LED to transmit digital information to a commodity camera. Conventional methods for object detection and identifications usually depends on the spatial appearance of an object or a marker, and their performance quickly deteriorate as the distance to the object increases, since the small number of image pixels occupied by the spatial pattern no longer has sufficient resolution. In this thesis, we take a different approach; instead of relying on only a spatial pattern, we also take advantage of a temporal pattern, i.e., the modulated light wave. To extend the range supported by the system, we leverage polarized light, which can hide flickers from human eyes while allowing the camera to reliably receive the transmission even at long distance. We built the PoliTag prototype using off-the-shelf components: retroreflector, polarizer, and a liquid crystal cell (LCC) in the Tag, and an electronic polarizer and a commodity camera in the Reader. The LCC of the Tag modulates the polarization angle of the emitted light to generate a polarization pattern that allows detection and identification of the Tag. The electronic polarizer allows Reader to calculate polarization information for each pixel. The polarization information allows us to reduce the region-of-interest (ROI) in each captured image to very small areas that contain only highly polarized pixels, reducing the required computation and improving the detection and identification performance. As a result, our implementation is capable of real-time processing at 100 frames per second. Evaluation results of our prototype show that the system has a 86% identification rate at 30 meters in challenging outdoor scenarios, while the Tag can have 16 different identifications and occupies only a 6 by 6 pixels area.