| With the development of artificial intelligence and machine learning, computer vision still plays an important role in the modern world. The information obtained from the outside by the eyes accounted for75%of all the information obtained by the human. This means how to capture the visual information quickly and accurately will make sense for the practical application. In this stage of study, visual tracking technology has a broad and in-depth research in the multidisciplinary field of machine monitoring, artificial intelligence and human-computer interaction. However, due to the actual motion of the target was often disturbed by the shelter, illumination changes, deformation, it is hard to track and locate the target.By analyzing the existing visual tracking algorithm, this paper mainly studied two factors which affected the tracking algorithm:target description model and target searching algorithm. To improve the accuracy of target location, this paper takes the MIL sampling and classification strategy. With the SIFT feature, this paper also uses the sparse matrix, the trivial template, and the improved weighted classifiers. The main contents of this paper are as follows:(1) Considering the feature based on simple gray information can not describe the target model accurately, this paper combines the improved SIFT feature and the trivial template to build an accurate model. After comparing some features which are similar to the SIFT feature, this paper selected the robust SIFT feature, and ensured the selected feature can distinguish between the backgrounds and objectives. To reduce the running time, this paper employed the sparse matrix to reduce the dimension of the feature, while retaining the structure of the original data. Combining the trivial template can also increase the robustness of this algorithm to noise.(2) By analyzing the different classification principles and the performance of diverse classifiers, this paper proposed the new weighted bayesian classifiers, with updating the weights by the sparse coding. The introduction of MIL can guarantee to select the more accurate positive bag and navigate to the optimum position.By comparing this algorithm with the other semi-supervised tracking algorithms, this algorithm can deal with the target drifting, shape variations and illumination changes. The experiment results show that the algorithm has good performance in the stability and the accuracy under the complex environment. |
