Research-based Non-rigid Motion Trajectory Based Recovery

Motion reconstruction is recover the3D structure and motion information from2D image feature points.It is an important problem in the field of Pattern Recognitionand Computer Vision. Now, it has been popular used in sorts of fields, such as thesimulation of cloth and textile、video surveillance、 security technology and so on.The reconstruction of non-rigid structure from motion is more difficult, because itsstructure is flexible and its shape is irregular.Nowadays, the reconstruction of non-rigid structure is always realized in shapespace. The main idea is that the3D non-rigid structure can be seen as linear weightedcombination of a series of shape bases. But the use of shape bases is confined, and weshould use different shape bases for different motion. For example, the reconstructionof walking and drinking should use different shape bases. So to realize motionreconstruction in shape space would cause large error. In order to solve this problem,the Non-rigid structure can be regarded as linear weighted combination of a series ofpredefined trajectory bases, so we can realize the motion reconstruction in trajectoryspace. Most existing non-rigid motion recovery is under orthogonal projection modelor weak perspective projection model, and these models belong to affine model, theyare the approximation of the real camera model. The orthogonal projection modelignore the depth and location information, the weak perspective projection modeldon’t ignore the depth information, but the model ignore the location information. Theassumption is valid only when the change of the object depth information doesn’tsignificantly. Or the assumption will cause large error and even make the recoveryinvalid.Aimed at these problems, the paper has carried on the related research, the mainwork is as follows:1、The paper implement the recovery of the non-rigid object3D structure andmotion from image sequences in trajectory space. Because the shape bases is confined,so on account of the time smoothness of the feature points in the3D space, thenon-rigid object can be seen as linear weighted combination of a series of trajectorybases. According to the duality of the shape basis and trajectory basis, we can changethe motion recovery from shape space to trajectory space. The paper analyze thedifferent type and number of the trajectory basis, and select appropriate type and number. Because the trajectory basis can be predefined, so the algorithm can reduceerror accumulation so as to enhance the stability and accuracy of the motion recovery.2、The paper proposes an iterative algorithm, with this algorithm and2D imagefeature sequences, we can realize the motion recovery of non-rigid object underperspective projection model. Most present algorithm are under weak perspectiveprojection model. It is the assumption of the true perspective projection model andignore the location information, that would cause large error. According to therelationship of the perspective projection and weak perspective projection, the paperweigh the feature matrix and use iterative algorithm to realize the approximate resultof motion recovery under perspective projection model. The experiment of the real2Dimage feature sequences show that the improved algorithm has better accuracy andstability.