Abstract

Inspired by recent advances in deep learning architectures, we propose a stratified architecture for visual tracking with deep learning. By using auxiliary natural images, we train a stacked denoising autoencoder offline to learn multi-layers image features that are more robust against variations. There are two levels in our approach: fine-grained tracking and coarse-grained tracking. In the first level, we track the object in each frame using a MOSSE correlation filter based approach, which is simple and efficient. A more complicated l1 minimization sparse representation based approach is used every several frames in the second level, where different features learned in different layers are used depending on the tracking effect. The tracking errors accumulated in the first level are corrected effectively. Combining the fine-grained tracking and the coarse-grained tracking, we achieve an efficient and robust visual tracking. Evaluation results on a number of sequences are given.

Paper Details