The Moving Objects Segmentation Based On Spatial-Temporal Information

In the paper, a moving objects segmentation algorithm based on spatial-temporal information is proposed to separate the moving objects from the background in a generic video sequence.At first, the background and the actuality of moving object segmentation is introduced, dissertate the moving object segmentation about the basic concept and characteristics, the generic classification of the moving object segmentation algorithms. Several typical algorithms is introduced according to the classification and the advantages and disadvantages are discussed at the same time.Secondly, the main theory and technique referred to the moving object segmentation algorithm based on spatial-temporal information proposed in this paper are introduced, including image smoothing, morphological processing, inter-frame differing in time domain and watershed transformation, the typical algorithms of which are also given, as well as their advantages and disadvantages. In this situation, the frame of the moving object segmentation algorithm based on spatial-temporal information is designed.In the temporal domain segmentation, the dual frame difference method is used here because of its simpleness and easy actualization. For this method, the performance of the segmentation in temporal domain depends on the threshold chosen. The moving area would probably be reduced if the threshold is set too high, while the background noise would increased if the threshold is set too low. In this situation, the Least-Half-Samples (LHS) [7] method is applied here to realize threshold processing. Meanwhile, how to fill the area of the initial segmentation result is also one of the emphases discussed in the temporal part of this algorithm. The dual scan method is used in the literature [7]. Good performance of area filling towards the protruding objects can be produced with this method, with which it can lead to area error-filling (Figure 3.6). In order to solve this problem, the area filling method based on morphologic is applied, which solved the problem efficiently. Through the segmentation in temporal domain, the initial segmentation mask is obtained, but because of the existence of the situation showed in Figure 2.3(b) , the segmentation mask is not precise, as there are pixels belong to the background on the edge zone of the mask. Therefore, the spatial segmentation is applied to make the edge more exact.In the spatial domain segmentation, the moving object segmentation mask is quite accurate, but it needs some modification to make the mask edge more precise , which put forward a higher require of the edge details, while the watershed algorithm is right the method to meet the require because it can be easily carried out and obtained close edge with one-pixel-width to get the accurate contour through segmentation, which is the reason why the watershed algorithm be used here. But there is a big problem of watershed algorithm, that is, over-segmentation. For solving the problem, a method is used to restrain the over-segmentation, that is, reducing the number of local small extremum regions through modifying the gradient image by minima imposition technique. At the same time, considering the characteristics of moving object segmentation, the temporal segmentation result is combined in the process of the spatial segmentation. Only the gradient image located in the area marked by the temporal segmentation mask is marked, so that the calculation is simplified. In other words, through the twice modification of the initial gradient image, it achieved the over-segmentation and simplified the calculation.Then, the conditions of the simulation experiments including hardware and software, and the source of the experiments object-YUV test sequence are introduced. For the moving objects segmentation algorithm based on spatial-temporal information proposed in this paper, the simulations are based on the YUV standard test sequence-akiyo and Mother_and_Daughter. Also, the whole simulation process and the result of the important steps are introduced. Through analysis of the simulation result, the experiments conclusion is given.At last, a summary of the works done in the paper is given, and the problems haven't been solved, as well as the research emphases and directions are proposed.In summary, there are five main work aspects carried on in the paper as follows:1. The moving object segmentation algorithms is researched and compared with an emphasis laid on the segmentation algorithm based on the spatial-temporal information of the video sequence.2. On the basis of present segmentation algorithm based on spatial-temporal information, a new moving object segmentation algorithm is proposed, which apply the dual frame difference method based on LHS threshold processing to the temporal segmentation, and watershed algorithm based on markers and segmentation result obtained in the temporal segmentation to the spatial segmentation.3. In the temporal segmentation, some improves are made on the basis of LHS threshold processing proposed in Literature [7], including applying adaptive median filter to pre-processing, and using the area filling method based on morphologic instead of dual scan referred in Literature [7] to avoid area error-filling efficiently.4. In the spatial segmentation, on the basis of initial watershed transform, a spatial watershed algorithm based on markers and temporal segmentation result is proposed. Firstly, the initial gradient image is modified for the first time with the temporal segmentation result, followed by the second modification with local small extremum markers. Then the watershed transformation is done so that the phenomena of over-segmentation is restrained and the calculations is simplified compared to usual watershed algorithm.5. The experiments is carried on the standard YUV testing sequence. Analyses are made on the simulation process of the algorithm and related matters proposed in the paper.Plentiful simulation experiments prove the feasibility and veracity of the moving objects segmentation algorithm based on spatial-temporal information proposed in the paper.