Siamese Matching Network Based on Robust Feature Representation for Object Tracking

Abstract

We propose a target tracking method based on siamese matching network for robust feature representation. Our network consists of region proposal layer, convolutional layer and long short-term memory layer. The method proposed in this paper takes advantage of deep learning in feature representation and the hierarchical structure of the convolutional network to extract different levels of information from different layers to obtain richer feature representation. The long short-term memory network is used to encode feature extracted by convolutional layer into a fixed vector, which can remember useful information to better capture the difference between images, so that the obtained feature vectors are more robust. The presented network matches the feature of target object with candidate region in current frame and returns the most similar region for tracking. We use external data sets for pre-training and the proposed method shows competitive performance on the standard tracking benchmarks.