以物理學為基礎之雨滴去除法應用於移動物件偵測 ; Physics-based Rain Removal for Moving Object Detection

NCU Institutional Repository > 資訊電機學院 > 資訊工程研究所 > 博碩士論文 > Item 987654321/44703

請使用永久網址來引用或連結此文件: http://ir.lib.ncu.edu.tw/handle/987654321/44703

題名:	以物理學為基礎之雨滴去除法應用於移動物件偵測;Physics-based Rain Removal for Moving Object Detection
作者:	陳冠志;Guan-jhih Chen
貢獻者:	資訊工程研究所
關鍵詞:	雨滴去除;移動物件;物件偵測;物理學;Object Detection;Moving Object;Rain Removal;Physics
日期:	2010-07-27
上傳時間:	2010-12-09 13:53:12 (UTC+8)
出版者:	國立中央大學
摘要:	在影片或監控系統下偵測移動物件時，常有物理雜訊干擾物件偵測。將雜訊去除以提高物件偵測識別率一直是物件偵測上一個重要的課題。在不同天氣下常有不同雜訊去除的議題必須克服：雨，由於它在空間上的隨機分佈與時間上的快速移動，是相當具有挑戰性的雜訊。這篇論文建立且改良幾位學者所提出偵測雨滴與背景估測的方法與相關策略，然而若只是根據雨滴在亮度上變化的性質，初步所偵測的像素位置並非完全屬於雨滴。為了將誤判的區域去除，我們使用正確率高的特徵分析，在多層感知類神經網路(Multilayer Perceptron Neural Network, MLP)的訓練下，將相關環境的特徵屬性建構出來。分析特徵之一是在 YCbCr通道下對每一個像素(Pixel)計算其於CbCr的變化量，此變化量大所對應的像素位置存在移動物件的可能性較高；特徵之二是Sobel邊緣偵測，通常有雨滴經過的區域其邊緣資訊會比沒有雨滴經過的較小；特徵之三則是像素的飽和度資訊，有雨區域存在一飽和度範圍區間。此分類的結果可用於判斷分析，除了初步去除誤判區域，運用於色彩變化相關性的特質，更可以將移動物所位於的誤判像素去除。我們的策略包含使用雨滴在時間與空間上的相關性，標示出雨滴所存在的區域並將這些去域範圍內的雨滴去除；在時空相關性較低的戶外條件下，仍可用類神經與色彩變化相關性來去除可能誤判的移動物。我們將此雨滴雜訊去除的架構應用於背景相減，找出正確的前景。將找出的前景做前景區域的正確率分析，所測試環境為正常光照下固定式攝影機的影片檔。實驗的結果將顯示我們的方法改良之前的幾篇研究並有效率的偵測出前景。實驗結果顯示此一架構的確有助於雨天天候下的前景偵測，我們得知雨滴去除雖然不會提升偵測正確率，但大幅降低偵測誤判率，在背景相減之前景物偵測仍相當有幫助。Physical noises, such as rain, frequently affect the detection performance of moving-objects when they present in a film or a monitoring system. The elimination of physical noises is thus a prerequisite to uplift the detection accuracy. There are different kinds of noises needed to be eliminated under different weather conditions. The elimination of rain drops, due to the random spatial distribution and fast motion in short time, is hence admitted as a challenging problem. In this thesis, we improve the detection method of rain drops and construct the estimation on the intensity of background. Extensive studies were conducted in analyzing these algorithms. However, not all of the detected pixels contain real rain drops if merely find them based on the property of intensity change. To remove the false alarm, we use three effective features in our work and construct the environmental parameters of features by MLPNN (Multilayer Perceptron Neural Network) training. One of the features is the changed value of color from Cb and Cr. The probability of a pixel containing moving objects is higher if the changed value is high. The second feature is the information obtained by the sobel edge detector. Generally, the lower value in edge will exist in the rain area. The last feature is the saturation value of a pixel. A range of saturation values exists in the rain area. The result of classified method can be used for making decisions on false alarm, or we can use it on the algorithm in constructing the color-correlation to map more false alarms. The color-correlation is very useful for removing the false alarm in moving objects. The proposed method is implemented on object detection to find real foreground. We analyze the detecting accuracy with and without the rain removal system in the normal illumination cameras or films. The results show that the proposed system is indeed helpful on moving-object detection in the rainy weather. It really reduces the false alarm although the accuracy is not raised with the rain removal system.
顯示於類別:	[資訊工程研究所] 博碩士論文

文件中的檔案:

檔案	描述	大小	格式	瀏覽次數
index.html		0Kb	HTML	845	檢視/開啟

在NCUIR中所有的資料項目都受到原著作權保護.

社群 sharing

資料載入中.....