Researches On Three-Dimesional Motion Tracking Method For Monocular Video Object

For a given video sequences,the goal of object tracking is to locate the tracking target in each frame.Object tracking as one of the most important research fields is widely used in virtual reality,industry control,military device,medical research,video surveillance,and intelligence traffic.3D pose estimation has become one of the hot research topics in the field of object tracking because it has the advantages of being low cost,simple and widely apply.The most employed approches in the monocular video sequences tracking are try to estimate the 2D motion of the object.Although,many techniques have been developed for vision-based 2D tracking,it still remains a challenging task because some key problems do not well addressed,such as: due to the absence of depth,2D motion tracking cannot used in some applications where the 3D pose of the target is required.Moreover,the depth value of an object will be lost when the object is projected onto 2D image plane,consequencely,3D motion reconstruction from 2D motion sequences is still a challenging task.Therefore,it is of great practical significiance to study 3D motion tracking from monocular video sequences.In this thesis,the 3D motion tracking method based on monocular video is thoroughly analyzed and studied.The thesis make research on cylinder model estimation,perspective invaniant point extraction,and 3D motion tracking based on reconstructed model.The contributions of this thesis summarized as follows:(1)An on-line Three-Dimensional Columnar Model of Target(TDCMT)estimation algorithm is presented in this thesis.Based on this,a 3D motion tracking is proposed.Firstly,the camera model is established which is determined by a scale factor.Secondly,an object function is created which related to the scale factor and the apperance model of the tracking object,and the Levenberg-Marquardt alogrithm is used to optimize the object function.Finally,a 3D motion tracking is proposed,in which the estimaed 3D object model is employed as the appearance model of the tracking object.As compared to the traditional hand made method,the proposed approach is more precise.(2)A Perspective Transformation Invaniant(PTI)key-point approach is proposed which can improve the performance of feature matching.The basic idea of the proposed algorithm is: estimating the truthly sample region of the feature descriptor by camera projection to eliminate the side affects which caused by the 3D viewpoint change.The implement of the proposed algorithm is: after estimated the object's pose in the prior frame,the sample region is determined by projecting the original sample region to the image plane based on the established camera model,to overcome the side affects which caused by the rapid motion,the estimated sample region is further refined by an optimization.The main novelty of the proposed approach include:firstly,A robust SIFT descriptor is proposed,which can mostly eliminate the side effects caused by perspective effect,so that the description is invariant to changes of 3D viewpoint.The gain in feature matching can improve the performance of tracking.Secondly,A Smoothness term is added to the pose estimation function,such a penalizer allows for outliers in the data(e.g.,due to noise,specularities,occlusions)and in the smoothness assumptions(due to motion discontinuities).Therefore,PTI can improve the prcise of matching,and increase the number of matched correspondences.(3)Based on the Reconstructed Three-Dimensional Model(RTDM),a 3D motion tracking approach is proposed.First,RTDM use first serval frames to reconstruct the 3D model of the tracking target based on the Structure From Motion(SFM)algorithm.To improve the precise of the estimated model,three effective steps are included:(a)background subtract;(b)frame test for model estimation;(c)dense 3D model estimation.Second,the pose of the object in other frames are tracked by a Extended Kalman Filter.As compared to the tradtional approach,RTDM is faster and more convenient because the model of the object employed is not required to be created by a 3D scaner or approximated by a geometry model.Besides,as the model estimated by computer vision,the precise of the estimated model is very high which lead to better tracking performance.