快速的移動陰影移除演算法與其在交通流量偵測上的應用

Abstract

近年來，越來越多智慧型影像監視系統被應用在提升人們的安全與生活品質，在大多數的這類系統中，前景物體擷取（Foreground object extraction）是一個非常重要而且基本的步驟，因為許多後續的處理與應用都是建立在前景物體上。然而移動陰影（Moving shadow）卻是影響前景物體擷取的一個關鍵因素。在戶外的環境下，光線被前景物體遮擋的時候便會產生陰影，而這些陰影常常會被錯誤地分類成前景區域，這樣的錯誤接著就會引起許多問題，像是物體定位會因為中心點偏移而出錯，而物體的外型邊線會變形。此外，如果兩個獨立的物體因為陰影而相連在一起，就可能會被判斷成只有一個前景物體。這些問題都會影響後續在追蹤、分類與辨識上的效能。 此外，許多的影像監控系統會偏好使用黑白攝影機，尤其是使用在戶外環境之下的系統。因為黑白攝影機會比彩色攝影機有較高的解析度，而且在低照度的情況下也會有較佳的影像品質。因此我們提出一個不需要使用彩色資訊的陰影移除演算法。藉由使用物體的邊線特徵（Edge feature），並且保持陰影區域內部的同質性（Homogeneous property），然後我們將物體邊線最外圍的部份去除掉，接著便可以得到非陰影的線條特徵。另外，我們也使用灰階的資訊建立出”變暗比率”（Darkening factor）的高斯模型，然後藉由這些模型來找出其他非陰影的特徵。接著合併這兩種非陰影的特徵，我們便可以去除陰影的影響且正確地框出前景物體的區域。 最後在車輛流量偵測的實驗當中，從三個測試影片所得到的數據裡可以看出我們的演算法可以提升整體4%~10%的正確率。此外，本論文所提出的移動陰影移除演算法在處理速度上平均每幀畫面只需要13.84毫秒，是相當有效率的。

In recent years, utilizing video processing to help for improving safety or human’s life has attracted great attention. Most of these application systems, foreground object extraction is a very fundamental step before further processing. However moving shadow is a critical influencing factor when extracting foreground object. In outdoor scene, moving shadow occurs when the light is blocked by moving object, and the shadow region is usually misclassified as foreground region. It would bring out a lot of problems. For example, shadow region may cause object localization problem, and shape deformation. Besides, if shadow region connects these objects, two or more independent objects would be treated as only one foreground object. All of these problems will degrade the performance of subsequent processing, like tracking, classification or recognition. In addition, some application systems prefer B/W (Black & White) camera rather than color camera especially in outdoor, because B/W camera have better resolution than color camera, and the sensing quality under low illumination condition is also better than color camera. Therefore, we propose a moving shadow removal algorithm without utilizing color information. We use the edge feature of object and keep the homogeneous property inside shadow region as much as possible. By eliminating the boundary edge of object, we can obtain the non-shadow edge feature. Additionally, we also utilize gray level information. We build a Gaussian darkening factor model for each gray level, and use these models to extract non-shadow feature. By integrating these two features, we can successfully detect the objects without including their shadow region. Finally, we take an experiment on vehicle counting. In our three test videos, the counting result can improve accuracy rate 4%~10% after using our shadow removal algorithm. The moving shadow removal algorithm proposed in this thesis has been successfully evaluated that the processing average time is 13.84 milliseconds per frame, and it is quite efficient.