| Summary: | 碩士 === 國立清華大學 === 資訊工程學系 === 104 === Foreground segmentation is a vital step for many high-level applications such as elderly surveillance, public safety, and tra c monitoring. While there are exten- sive methods that have been proposed for foreground segmentation/background subtraction, most of them assume that cameras are stationary which means they treat each pixel individually. With this assumption, they are unable to handle movements caused by the moving camera. It is because each pixel represents var- ious of position in the moving camera. Therefore, false detections increase due to the lack of alignment between observed pixels and background models. Although there are few methods are proposed to address the alignment problem, they do not consider the impact of registration errors and suddenly large movements be- tween consecutive frames. Most of them are also unable to detect subtle changes in camou age objects because they usually use only color intensities to model the background. Besides, in classi cation step, most of them use a global threshold which is unable to present the various behaviors of background scenes such as waving trees. In this paper, we propose a robust hybrid-sample-based foreground segmentation method for moving cameras to address these problems, especially on pan-tilt-zoom cameras. First, we propose a motion clustering registration to re- duce the impact of registration errors. We estimate a homography matrix between two consecutive frames. Thus, movements of pixels can be estimated by using predicted homography transform, and a motion clustering registration re nement is adopted to minimize the impact of registration errors. Second, a frame-level reinitialization scheme is proposed to solve a suddenly large movement between consecutive frames. Third, we propose a hybrid-sample-based background model- ing technique that each pixel is modeled by not only a color intensity value but also texture information. A novel robust binary descriptor is presented for the background modeling. This allows us to easily detect camou age foreground ob- jects which have the similar color to the background scene. Last, in order to deal with dynamic background, we adopt pixel-level feedback schemes to dynamically and locally control the sensitivity and the adaptation speed of the background model. We evaluate the proposed method with the ChangeDetection.NET 2014 dataset. Experimental results show our detection results are more robust espe- cially in camou age foreground regions. The shape of detection results is also more accuracy. The motion clustering registration can eliminate most of noises caused by registration errors. The proposed method is 10 percent better than other state-of-the-art algorithms in terms of overall F-score of panning sequences, and it also achieves the highest F-score in camera jitter scenarios.
|
|---|
