Research Of RGBD Video Segmentation Based On Graph

Video segmentation is one of the most important aspects in various image-based research and applications,and it's also the basis of many vision algorithms.The quality of video segmentation directly affects the performance and efficiency of those algorithms.With the development of various visual devices,it becomes easier than ever to obtain RGBD data,research and applications based on RGBD data are more and more.Therefore,video segmentation for RGBD data has become one of the hot spots in computer science.Video segmentation algorithms could be grouped into several types,and video segmentation based on graph is an important method among these algorithms.In graph-based video segmentation algorithm,the main steps can mostly be divided into two phases:pre-segmentation and merging.In the pre-segmentation process,video segmentation should be precise enough and match the facts.In the merging process,most of the algorithms are not adaptable to the specific scene,and often require manual intervention to best segment the specific scene.Last but not least,region matching is also a hot issue in image post-processing,it matches the results between two neighboring segmentations,which ensures the consistency in the whole process.This paper has conducted in-depth research on these three aspects.The main achievements are as follows:1.Based on a pre-segmentation algorithm of Hickson,a new pre-segmentation process is added and implemented based on normal vector and curvature,which make the results of pre-segmentation more precise,benefit the following merging process.Experiment verifies that dividing the pre-segmentation into several sub-process based on different features is better than methods based on combined features in performance.2.A region merging algorithm based on affinity propagation clustering is proposed and implemented,which can adaptively get the number of clustering center,avoid the problem that the performance is influenced by the scenarios themselves and needs manual intervention to set the threshold in hierarchical merging.It effectively improves the generality.3.A neighboring frames matching algorithm based on KM algorithm is proposed and implemented.We take advantage of the characteristic which KM algorithm could distinguish similar edge weights,improving the accuracy of image matching.Particularly,when there are several regions with similar region characteristic in the situation,the accuracy of this algorithm could be much better.