- 學位論文
夜間近紅外線影像之人物偵測
Human Detection in Nighttime Near-InfraRed Video
摘要
在本篇論文中，我們針對夜間近紅外線影像作處理，期望能夠強化近紅外線影像之視訊品質，並對前景人物作偵測。我們使用時空域雙向濾波器為基礎，依據時間、空間鄰近像素點之關係進行各個像素點之雜訊濾除，此關係包括距離及相似度。接著，我們使用高斯混合模型進行人物偵測，並對夜間感應式照明設施環境及其導致的光影現象進行額外的處理。在此我們提出雙模式高斯混合模型，即分開建構開、關燈環境之背景高斯模式，並藉由感應式照明設施的開、關燈切換之事件偵測進行背景高斯模式的切換。由於開燈環境下通常會強化影子效應，我們藉由掃描線搜尋法搜尋人物的立足點並利用此資訊濾除人物之影子。因此本論文所提出的方法，能夠提供偵測前景人物所得之外接矩形資訊，將有助於應用近紅外線影像之夜間監控系統。
並列摘要
In this thesis, image processing techniques are applied to the analysis of near-infrared videos. The goal is to detect human activities in the videos. We implement a method to enhance the video quality, a spatial-temporal bilateral filter which filters out each pixel’s noise based on the spatial and temporal relationship, including distance and similarity, with its neighbors. Subsequently, the Gaussian mixture model （GMM） is used to construct background model and to perform foreground detection. Additionally, we pay attention to sensing lighting equipments used in nighttime environment because of its illumination and shadowing phenomena. Accordingly, a two-mode GMM is proposed which separately constructs background GMM for different lighting conditions and switches GMM modes by event detection. In order to cope with excessive shadowing phenomenon, an efficient way of searching footholds by using scan-lines is proposed to remove human shadows. The proposed approach will provide bounding boxes information of human regions as detection results which will be very helpful for a nighttime surveillance system based on near-infrared videos.