A practical vision system for the detection of moving objects

The main goal of this thesis is to review and offer robust and efficient algorithms for the detection (or the segmentation) of foreground objects in indoor and outdoor scenes using colour image sequences captured by a stationary camera. For this purpose, the block diagram of a simple vision system is offered in Chapter 2. First this block diagram gives the idea of a precise order of blocks and their tasks, which should be performed to detect moving foreground objects. Second, a check mark () on the top right corner of a block indicates that this thesis contains a review of the most recent algorithms and/or some relevant research about it. In many computer vision applications, segmenting and extraction of moving objects in video sequences is an essential task. Background subtraction has been widely used for this purpose as the first step. In this work, a review of the efficiency of a number of important background subtraction and modelling algorithms, along with their major features, are presented. In addition, two background approaches are offered. The first approach is a Pixel-based technique whereas the second one works at object level. For each approach, three algorithms are presented. They are called Selective Update Using Non-Foreground Pixels of the Input Image , Selective Update Using Temporal Averaging and Selective Update Using Temporal Median , respectively in this thesis. The first approach has some deficiencies, which makes it incapable to produce a correct dynamic background. Three methods of the second approach use an invariant colour filter and a suitable motion tracking technique, which selectively exclude foreground objects (or blobs) from the background frames. The difference between the three algorithms of the second approach is in updating process of the background pixels. It is shown that the Selective Update Using Temporal Median method produces the correct background image for each input frame. Representing foreground regions using their boundaries is also an important task. Thus, an appropriate RLE contour tracing algorithm has been implemented for this purpose. However, after the thresholding process, the boundaries of foreground regions often have jagged appearances. Thus, foreground regions may not correctly be recognised reliably due to their corrupted boundaries. A very efficient boundary smoothing method based on the RLE data is proposed in Chapter 7. It just smoothes the external and internal boundaries of foreground objects and does not distort the silhouettes of foreground objects. As a result, it is very fast and does not blur the image. Finally, the goal of this thesis has been presenting simple, practical and efficient algorithms with little constraints which can run in real time.