No Calibration More Adaptive Angle Tracking System

In a multi-view video tracking system, optimization and correction can be gainedthrough information exchange with other views to achieve better performance thansingle-view system. But this achievement could be got only when geometryrelationship between views is known to project positions from one perspectivecoordinate to another. Then, selection within different methods of using geometryrelationship is an important part of system design, and calibration, which meanssolving process of geometry relationship, becomes a main limiting factor in bothdeployment and operation of multi-view tracking systems. Homography is afrequently used geometry relationship which can map one-by-one correspondencebetween two perspective coordinates by restricting that tracked objects always moveon ground plain. Generally, control points needed to estimate homography isobtained by manually selection or placing specially marks on ground plain.In this paper, our finial target is to combine processes of tracking and estimationof homography together to avoid a pre-calibration process, and to launch a furtheroptimization on multi-view information fusion framework, making it adaptive to alltracking error, projection error and geometry relationship. This target is achievedwith three steps: perturbation analysis of homography estimation, selection methodof control points from tracking result and adaptive framework for multi-viewinformation fusion.Perturbation analysis of homography estimation is a quantitative analysis onprojection error. In the process of homography estimation, observation noise oncontrol points would cause an error in estimated projection matrix, which wouldfurther cause the projection error. In this paper, we analyze the quantitativerelationship between projection error and each observation noise. Then, with thisresult, we could both add projection error level to adaptive frame by estimating itand select proper control points by minimizing it.Selection of proper control points is the key of estimating homography withtracking result. In this paper, we start with the result of perturbation analysis andlaunch a further study on how a control point’s accuracy and position affectprojection error to approximate an unsolvable optimization problem. A progressivemethod and condition function is provided, which proved to be both effective andefficiency. An adaptive framework for multi-view information fusion based on GMM(Gaussian Mixture Model) is achieved, which includes all tracking error, projectionerror and mutual information level between views. By this framework； estimationprocess is combined this tracking process and fusion performance rs improved.