Researches On The Sparse Representation Based Object Tracking Algorithms

With the improvement of computer computing ability and the closer relationship between computer vision technology and people's life,computer vision research has gained more and more attention from researchers of the world.Object tracking technology is one of the hotspots in the field of computer vision.It is widely used in the field of guidance technology in military equipment,video surveillance in urban safety,intelligent transportation in traffic system and somatosensory games in game entertainment.Although the current object tracking technology has been widely used in many fields and the propose of a large number of algorithms makes the research of the object tracking technology achieve an impressive progress,but because of the actual scene consists of a variety of complex factors,such as background clutter,variantion in light,motion blur,and occlusion,it is still of great significance to design a robust and real-time object tracking algorithm.Therefore,based on the research of object tracking technology,this paper tries to solve some key problems in object tracking,and concentrates on the improvement of the appearance model of the object,the following innovations are obtained:1)a sparse representation tracking algorithm based on residual matrix estimation is proposed.At present,Most of existing tracking algorithms based on sparse representation use the sparse representation coefficients to construct the representation model,and the constructed observation model only uses the representation coefficients as the evaluation of each candidate particle,ignoring the residuals.In this paper,the residual matrix of the object is quantized as a residual vector and introduced into the sparse representation model.A sparse representation model containing residual vectors is constructed.The model uses the L1 norm to constrain the representation coefficient and the corresponding residual vector and use the L2 norm to constraint the distance between the reconstruction and the object.Secondly,in order to solve the representation coefficients model and the corresponding representation residual vector in each candidate particle representation,this paper adopts a method of cyclic iteration,which fixes one of the two variables separately and solves the other.Through this way,the approximate optimal solution of the coefficient and the residual vector can be obtained.Then we use the acquired representation coefficient and the residual vector to construct the observation model of each particle,and select the object of the current frame from the candidate particles according to the model score.Finally,in order to make the algorithm have a long and stable tracking effect,this thesis computes the distance between the dictionary template and the selected object as the template score,and divides the update step into three categories according to the score of the template.The experimental results show that the sparse representation tracking algorithm based on residual matrix estimation proposed in this paper has better robustness compared with the other algorithms.2)an improved residual matrix estimation algorithm for sparse representation is proposed.The sparse representation tracking algorithm based on the residual matrix estimation is less robust to the occlusion problem and therefore does not apply to some scenes with severe occlusion.Therefore,based on the sparse representation of the residual matrix estimation algorithm,the representation model and the observation model are improved,which makes the algorithm improve the processing ability of the occlusion problem.Firstly,the optimization model of residual matrix extimation algorithm is still used to estimate the representation coefficients and the corresponding residual vector of each candidate particle.When the residual vector of each particle is obtained,The algorithms filters the residual vector in the current frame by the residual vector the selected partilce in the previous frame,so that the residual vectors between the current and previous frames have consistency at the nonzero element position.Secondly,this paper uses the distance of the reconstruction of each candidate particle and the residual vector between successive two frames as basis of the particle observation model,and then selects the tracking result of the current frame according to the score of the model,In this paper,the frequency of each template denotes how many times the template is used the to represent all candiate particles,and then the algorithm use the frequency as as one of the scoring factors.Anothe scoring factor is the distance between the dictionary template and the selected object in the tracking algorithm.The construction of the final score formula consists of above two factors.Then according to the sort order of the score of each template,the algorithm have four different situations to update the dictionary.The experimental results show that the improved algorithm proposed in this paper has improved the processing ability of the occlusion problem compared with the original residual matrix estimation algorithm,and the algorithm has better robustness compared with the other algorithms.3)a sparse representation tracking algorithm based on PCA subspace is proposed.Subspace tracking algorithm is more robust to illumination,pose variation and so on,but it is more sensitive to occlusion and other problems.On the constrast,the sparse representation is more robust to the occlusion problem.Therefore,this paper proposes a sparse representation tracking algorithm based on PCA subspace.Firstly,the algorithm calculates the base vector of the object in the PCA subspace,and uses the spatial basis vector,the spatial mean value and the representation vector of the residual vector to represent the object.The L1 norm is used to constrain the residual vector and the representation coefficient of the object over the spatial basis vector.And the L2 norm constraints the error between the reconstruction and the object.Secondly,in order to obtain the representation coefficient and the residual vector,the algorithm uses the method of cyclic iteration to decompose the complex representation model into twostep solution.In this way,we can effectively solve the approximate optimal solution of the representation coefficient and the residual vector.Then,in the particle evaluation step,the algorithm constructs the observation model of each particle by using the reconstruction error of each particle and the corresponding L1 norm function value as the particle score factors.Finally,in order to ensure the dictionary's adaptation to the object,this algorithm determines a variety of update cases based on the comparison among nonzero elements ratio of the residual vector and the pre-set thresholds.Then the object results of certain tracking frames are collected to learn the new subspace basis vector and spatial mean value by incremental principal component analysis.The experimental results show that the proposed algorithm can adapt to the problems such as illumination,occlusion,pose variation and background clutter during the tracking process.Compared with other algorithms,the algorithm has better robustness.