Convolutional Neural Networks (CNNs) had become a powerful model for solving many problems. In this paper, a novel visual tracking algorithm in large scale video based a trained CNN is proposed. The algorithm can track the trajectory of a moving object in a video with complex background quite precisely. Different from the most existing algorithms, we offline pre-trained a CNN through massive images data to obtain generic image features, which can be used the online tracking process. The trained CNN consists of shared layers and multiple branches of domain-specific layers, each branch is used for classification to identify target in each domain. When tracking a moving object in a new video sequence, a new network by combining the shared layers in the pre-trained CNN with a new classification layer is constructed. Experiment results show that performance of the proposed algorithm is excellent for some representative tracking benchmarks.
展开▼
