A neural network approach to visual tracking

Fully Convolution Networks (FCNs) have been shown to be effective in semantic segmentation through fine-tuning classification networks on segmentation data. In this paper, we present that FCNs can be further fine-tuned on target-background images in order to solve visual tracking problems. Pixel level models (FCNs) trained on segmentation data are superior to class level models (e.g. VGG net and GoogLeNet) in visual tracking tasks due to their powerful ability in discriminating between objects and background. Our work is based on a FCN network structure. The result is achieved by first fine-tuning the first image of a sequence and then the tracking and updating processes are conducted through classical forward and backward processes of neural networks. The proposed model achieves high precision and tracking success rates in online object tracking benchmark (OTB) data. It indicates our approach is competitive to state-of-the-art approaches as well.