To improve the tracking accuracy and the robustness of object tracking by exploiting discriminative features of a tracking target effectively, an object tracking algorithm is proposed in the framework of particle filter tracking based on the feature selection and temporal consistency sparse appearance model. Firstly,some positive templates,negative templates,and candidate targets are sampled, and their corresponding features are selected according to the feature selection model. The redundant interferential information is deleted, and the key feature information is obtained. Secondly, a multi-task sparse representation model containing a temporal consistency regular term is established via the features of positive templates,negative templates, and candidate targets. It induces more candidate targets to have sparse representation similarities with the previous tracking results. Thirdly,the multi-task sparse representation model is solved to gain the discriminative sparse similarity map, and the discriminative score is obtained for each candidate target. Finally, the positive templates and the negative templates are updated according to the tracking results. Experiments demonstrate that the proposed tracking algorithm produces better accuracy than some tracking methods,even under the complex environments.

The rough k-means algorithm and its derivatives focus on the description of data objects in uncertain boundary areas. However, the influence of imbalanced sizes between clusters on the clustering result is ignored. The interval type-2 fuzzy measure is introduced in this paper for measuring the boundary objects and an improved rough K-means clustering algorithm is developed. Firstly, the membership degree interval of the boundary object is calculated according to the data distribution of clusters and thus the spatial distribution of clusters is described. Then, the data sample size of the cluster is taken into account to adaptively adjust the influence coefficient of boundary objects on overlapping clusters. The experimental results on both synthetic and UCI datasets show that the adverse impact of the boundary objects on the means iterative calculations of small sample size clusters is mitigated and the clustering accuracy is improved.