This dissertation deals with active contour model based on object model. Active contour model is a deformable contour that can converge on edges of an object in an image for obtaining information related with the object. However, in noisy environment, a contour may converge edges of unexpected objects because conventional active contour model uses only edge information. In order to solve the poor convergence problem, this dissertation proposes a new active contour model using three-dimensional object model in noisy environment.
In this dissertation a contour on two-dimensional image is assumed as the projection of three-dimensional rigid body object that moves freely in three-dimensional space. Therefore, the contour is deformed with restriction imposed by 11 variables that form projection matrix that represents relation between three-dimensional object and two-dimensional contour. This dissertation also proposes virtual electric field in order that a contour converges on edges of an object in an image. Virtual electric field is a vector field that is created by virtual charges. The intensities of edge image computed by edge operator such as sobel, prewitt, etc. are considered as the virtual charges from which virtual electric field is generated. Under virtual electric field, each node of a contour with positive virtual charge is forced toward the edge of an object. From the converged nodes of a contour by virtual electric field and three-dimensional object model, the projection matrix is obtained and a contour is re-projected by the obtained projection matrix and three-dimensional model for restricting deformation. Also camera intrinsic parameters and pose of the object are recovered from the obtained projection matrix.
For fast computation, the formula for virtual electric field is represented as the convolution of an edge image and pre-defined mask. In modern computing technology, the convolution is calculated rapidly by dedicated hardware such as DSP and MMX. Simulations and experiments show that the proposed algorithm operates at 10 frame/sec on a standard hardware and a contour can converge on the edge of the target object even under noisy environment.
