Visual Object Tracking via Deep Neural Network

Visual tracking is a fundamental research problem in computer vision field. In this paper, we propose an approach to incorporate visual prior into visual object tracking via deep neural network. Visual prior knowledge is expressed as the parameters of a stacked denoising autoencoder, which is trained from a large collection of natural images. By utilizing natural images, we can obtain generic image features which are more robust against variations. Then we design a classifier for tracking using the same structure as the stacked denoising autoencoder, tracking is then carried out under a particle filter framework by determining the current target's location and updating the parameters. In addition, in order to alleviate the computational burden caused by deep structure, an adaptive updating mechanism is proposed. As a result, we apply a general-to-special strategy for our stacked denoising autoencoder tracker (SDAT), the learned visual prior provides a reasonable initial value for parameters of the neural network, and the deep structure of our tracker is robust to appearance variations. Experiments over 50 challenging videos indicate the effectiveness and robustness of our tracker, and the resulting tracker is outstanding especially against variations with the existing state-of-the-art methods.