Research On Template Matching And Multi Feature Fusion For Visual Object Tracking

Object visual tracking is an important branch in the field of computer vision.With the increasing difficulty of the tracking task and the change of the scene,the tracking process is facing great challenges.According to the limited sample construction suitable matching template set and how to solve the single feature in the process of tracking instability are important research direction in the field of target tracking,and they are the main contents of this paper.The selection of samples plays an important role in the process of target tracking.In order to select samples better,we propose an online sample selection algorithm based on non-negative least squares.As for the discriminative tracking method,a set of templates is usually constructed to measure the matching degree between the candidate samples and the template by scoring function.The training samples are too small in the template,so the model can not well represent the real situation of the target.On the contrary,the model is too complex to calculate.The non-negative least squares online sample selection algorithm can keep the template set as a linear combination of a representative sample based on the established objective function,through the coefficient of constraint model to selection the representative samples,so as to solve the problem of sample selection.At the same time,the objective function can be transformed into the least squares prob lem.Then the model can solved by standard nonnegative least squares method.Aiming at the problem that the representation ability of the single feature to the target attribute can not deal with the complex scene change,a target tracking algorithm based on multi feature fusion has proposed.The algorithm uses the weighted average of the largest response sites with different features to determine the center of the target.Multi feature fusion can make use of the complementarity between different features to improve the adaptability of the complex scene,and make up for the disadvantage of single feature tracking.At the same time,the weight of the response points of each feature is determined by the corresponding maximum response value,which can give full play to the role of the response site and enhance the robustness of the algorithm.We evaluated the performance of our methods on two datasets,Object Tracking Benchmark(OTB)2013 and Object Tracking Benchmark(OTB)2015.The two test sets contains more than 30 kinds of target tracking methods and 100 video test sequences.The experimental results show that the proposed methods are superior to the classical algorithm,and has the same performance with the algorithms which has the leading level.