Object-based video analysis describes the contents of the video focusing on the information about the objects, such as shape, color, intensity and gesture, and etc. Object-based technologies take an important role in video analysis. Moving object segmentation is the heart of the object-based technologies and can be applied to various applications such as object recognition, object tracking, object-based video indexing and object-based video coding. So, it has been studied for the last few decades in computer vision and image coding area. As the result, many algorithms and systems have been proposed and developed. In spite of these efforts, however, it is the one of the most difficult problem yet.
The best-known mechanism for moving object segmentation is the Human Visual System (HVS). Moving object segmentation is the high level analysis. The performance evaluation is answered by the person. So, in the moving object segmentation process, it is important to extract the object boundary coinciding with HVS because HVS has high sensitivity about motion and edge.
In this paper, we present a motion segmentation algorithm that adaptively combine color and motion based on the sensitivity of the human visual system. HVS (Human visual system) is very sensitive to moving object boundary. In other words, edge and motion are the main features that HVS perceive intensively. So, it is very important to obtain accurate boundary of moving object in image sequence. Proposed algorithm is composed of three parts: color segmentation, motion analysis, and region refinement and merging part. In the color segmentation phase, K-Means algorithm is used to cluster the color bins that have weights in consideration of human color perception. The global and local motion estimation are performed in parallel with color analysis. As the result, boundary and motion of each region are obtained. After that, Bayesian clustering using color and motion is proposed to obtain more accurate boundary for each region. By performing Bayesian clustering, region boundaries are refined although the color contrast between objects and background is low. Also, accurate region motion is obtained. In the final stage, regions are merged taking into account their motion. The experimental results of the proposed algorithm show the accurate moving object boundary coinciding with the boundary that HVS perceive.
