Level Set Based Online Visual Tracking via Convolutional Neural Network

In this paper, we propose a level set tracking algorithm, which integrates the information of the original frame and the confidence predicted by the deep feature based detector. First, we extract features from convolutional neural network and select part of them to avoid redundancy. Secondly, the features are used to generate a confidence map of the tracked object through the detector. And then the confidence along with the original frame is applied in level set model to acquire the segmentation result. We introduce an outlier rejection scheme to further improve the result. Finally, updating is employed to the detector to adapt to the changes in the video. One important contribution of our work is to use the deep features in confidence prediction, particularly the usage of low-level features in the neural network. Experimental results show that our model delivers a better performance than the state-of-the-art on a series of challenging videos.