| Summary: | 碩士 === 義守大學 === 資訊工程學系碩士班 === 99 === In order to detect fire accident in the open space, the method is used by detecting smoke in the video. In this thesis, target tracking method and the properties analysis of smoke are used to confirm whether there is smoke in the video. First, moving objects in video are detected by using Gaussian Mixture Model algorithm, then labeling disconnected areas as different parts. After labeling, the location of current image components and previous image components are compared to confirm whether the same components, then the smoke features of each component are extracted separately and estimate the motion orientation. In smoke feature extraction, the chrominance of the component is checked and the component will be marked if it fits the condition. In motion orientation estimation, time and cost are saved by using integral image. Moreover, by calculating UMR (Upward Motion Ratio) the components of non-smoke motion like will be excluded. Finally, the results of all steps are combined to confirm whether smoke presences in the video. According to the experiment, thirteen videos that have different complexity background were used, and the smoke detection rate was 92%, while in the false positive rate was only 7%. Compared to the program results of based on wavelet transform, the method in this thesis has greatly reduced the false positive rate and the interference caused by non-smoke objects in video.
|
|---|
