Multi-task Visual Tracking Using Composite Sparse Model

Despite the numerous algorithms proposed in the literature, objecttracking remains a challenging problem due to the complex real-lifescenario. Tracking algorithms that can steadily track the object undercomplex scenario is still urgently needed.Under the framework of particle filter, we propose a trackingalgorithm based on sparse representation and multi-task learning.Compared with traditional sparse coding based tracking methods, our workis advantageous in three ways:Firstly, in order to reconstruct the particles accurately, we see thesparse representation process as a multi-task learning problem, andpropose the composite sparse model(CSM) to improve the learningefficiency. The CSM learn the target parameter coefficient matrix andtrivial coefficient matrix separately. The trivial templates are incorporatedinto the model to reconstruct partial occlusions which can occur anywhererandomly. We regularize the trivial parameter matrix withl1,1norm to obtainelement-wise sparsity. The target templates are incorporated to reconstructthe appearance of the particle as a whole. And because the particles usuallylay near the region of the target, they are similar in appearance to eachother. This similarity leading to the fact that the particles can be seen asbeing reconstructed with common templates that most particles share andindividual particles that are specific to each certain particle. The CSM seesthe target parameter matrix as the addition of the common templatesparameter matrix and the individual templates parameter matrix. Weregularize the common templates parameter matrix withl1, norm to obtaingroup-wise sparsity, and regularize the individual templates parameter matrix to obtain element-wise sparsity. The CSM not only improve thetracker’s ability to reconstruct occlusions, but also undermines theco-relation between particles, leading to the declining of the reconstructionerror of sparse representation.Secondly, to comprehensively record the appearance change of thetarget, we propose a multi-threshold template updating scheme. In onesingle target templates dictionary, the target templates are updated withdifferent frequency: the templates with high updating frequency cancapture the current appearance change of the target, the templates with lowupdating frequency can record the history of the target appearance. In thisway, the multi-threshold updating scheme can make the target dictionaryrecord the target appearance comprehensively. This makes the tracker morerobust to the change of target appearance.At last, we propose to use ADMM(alternating direction method ofmultipliers) to optimize the CSM. Compared with the convex optimizationmethod other sparse coding based trackers use, ADMM can convergefaster, leading to fast tracking speed of the CST(Composite SparseTracker).We compare our CST with seven state-of-the-art trackers, and theexperiment results show that the CST can accurately tracks the target underconditions of partial occlusions and changing target appearance.