01-03-2010, 03:33 PM
Abstract
In this paper, a novel moving object modeling suitable for multicamera correspondence is introduced. Taking into consideration the color and motion features of foreground objects in each independent video stream, our method segments the existing moving objects and constructs a graph-based structure to maintain the relational information of each segment. Using such graph structures reduces our correspondence problem to a subgraph optimal isomorphism problem. The proposed method is robust against various resolutions and orientations of objects at each view. Our system uses the fuzzy logic to employ a human-like color perception in its decision making stage in order to handle color inconstancy which is a common problem in multiview systems. The computational cost of the proposed method is made low to be applied in real-time applications. Also, it can solve the partial occlusion problem more precisely than the Meanshif occlusion solver by 15.7%.
Presented By
Amir Hossein Khalili1 and Shohreh Kasaei2
1 Sharif University of Technology, Tehran,
2 Sharif University of Technology, Tehran,
1. Introduction
The increasing demand for analyzing moving objects behavior in wide area has heightened the need for robust modeling and correspondence of the moving objects in scene. The purpose of object modeling is taking out some features from images of an object so that the selected features are stable and reliable and could exactly discriminate the target in consequence projections. This task is difficult because an object in monitoring systems can move fast and unpredictably, can appear in a variety of poses and colors, and are often surrounded by clutter. For applications such as outdoor video surveillance, where the entire scene is not coverable with a single camera, a distributed camera network should be implemented. In multiview framework, differences in camera directions, distances to target, illumination situations and quality of capturing cause single object is looked differently in each field of view. Changes in size, view direction, luminance and color values of objects are some artifacts which make the problem more challenging. For that reason, special object model is needed to maintain correspondence of two-dimensional projections of an object seen in different views. Wide variety of methods has been reported to model moving objects in multiview applications, but a few of them handle color variation and occlusions effectively without employing high-level reasoning procedure and predefined model of target objects. In our experience the low level image features play a crucial role. The proposed method for object modeling aims to segment moving object efficiently according to its color and motion features and accommodate human-like color perception to deal with uncertainty in observed colors. The proposed method constructs a region adjacency graph for each moving object perspective to book relational state of each segment and tries to reduce moving object correspondence problem to a subgraph optimal isomorphism problem. The proposed method doesnâ„¢t need any high-level definition of moving objects and could deal with partial occlusions efficiently. The rest of paper is organized as follows. In section 2 an overview on related previous works is given. In section 3 different steps of the proposed algorithm are introduced in detail. The experimental results are shown in section 4. Finally, section 5 concludes the paper.
read full report
http://ipl.ce.sharif.edu/Papers/Csicc_08_E_AHK.pdf
In this paper, a novel moving object modeling suitable for multicamera correspondence is introduced. Taking into consideration the color and motion features of foreground objects in each independent video stream, our method segments the existing moving objects and constructs a graph-based structure to maintain the relational information of each segment. Using such graph structures reduces our correspondence problem to a subgraph optimal isomorphism problem. The proposed method is robust against various resolutions and orientations of objects at each view. Our system uses the fuzzy logic to employ a human-like color perception in its decision making stage in order to handle color inconstancy which is a common problem in multiview systems. The computational cost of the proposed method is made low to be applied in real-time applications. Also, it can solve the partial occlusion problem more precisely than the Meanshif occlusion solver by 15.7%.
Presented By
Amir Hossein Khalili1 and Shohreh Kasaei2
1 Sharif University of Technology, Tehran,
2 Sharif University of Technology, Tehran,
1. Introduction
The increasing demand for analyzing moving objects behavior in wide area has heightened the need for robust modeling and correspondence of the moving objects in scene. The purpose of object modeling is taking out some features from images of an object so that the selected features are stable and reliable and could exactly discriminate the target in consequence projections. This task is difficult because an object in monitoring systems can move fast and unpredictably, can appear in a variety of poses and colors, and are often surrounded by clutter. For applications such as outdoor video surveillance, where the entire scene is not coverable with a single camera, a distributed camera network should be implemented. In multiview framework, differences in camera directions, distances to target, illumination situations and quality of capturing cause single object is looked differently in each field of view. Changes in size, view direction, luminance and color values of objects are some artifacts which make the problem more challenging. For that reason, special object model is needed to maintain correspondence of two-dimensional projections of an object seen in different views. Wide variety of methods has been reported to model moving objects in multiview applications, but a few of them handle color variation and occlusions effectively without employing high-level reasoning procedure and predefined model of target objects. In our experience the low level image features play a crucial role. The proposed method for object modeling aims to segment moving object efficiently according to its color and motion features and accommodate human-like color perception to deal with uncertainty in observed colors. The proposed method constructs a region adjacency graph for each moving object perspective to book relational state of each segment and tries to reduce moving object correspondence problem to a subgraph optimal isomorphism problem. The proposed method doesnâ„¢t need any high-level definition of moving objects and could deal with partial occlusions efficiently. The rest of paper is organized as follows. In section 2 an overview on related previous works is given. In section 3 different steps of the proposed algorithm are introduced in detail. The experimental results are shown in section 4. Finally, section 5 concludes the paper.
read full report
http://ipl.ce.sharif.edu/Papers/Csicc_08_E_AHK.pdf