PROJECT TITLE :

Video Tracking Using Learned Hierarchical Features - 2015

ABSTRACT:

During this paper, we propose an approach to be told hierarchical options for visual object tracking. Initial, we tend to offline learn features robust to various motion patterns from auxiliary video sequences. The hierarchical features are learned via a two-layer convolutional neural network. Embedding the temporal slowness constraint in the stacked architecture makes the learned options strong to sophisticated motion transformations, that is vital for visual object tracking. Then, given a target video sequence, we propose a website adaptation module to on-line adapt the pre-learned features in keeping with the particular target object. The adaptation is conducted in both layers of the deep feature learning module so as to incorporate appearance info of the particular target object. Therefore, the learned hierarchical options will be strong to both complicated motion transformations and look changes of target objects. We have a tendency to integrate our feature learning algorithm into three tracking strategies. Experimental results demonstrate that vital improvement will be achieved using our learned hierarchical options, especially on video sequences with sophisticated motion transformations.