Multi-feature fusion Siamese Network for Real-Time Object Tracking

Abstract

In the multilayer neural network, the features of the low-level layers are of high resolution, which is suitable for positioning the object, while the features of the high-level layers are of rich semantics features which are suitable for the classifying the object. In order to utilize the advantage of high-level features and low-level features, we introduce a densely connected network called DSiamFc(Densely Connected Siamese Networks). Not only the low-level features and high-level features are fully integrated, but also this connection method can provide better parameter adjustment for the whole network during off-line training for the end-to-end object tracking network. The effectiveness of our proposed network is demonstrated by analyzing the backpropagation of gradient flow. Our algorithm is able to achieve real-time, and in the OTB-2013/50/100 benchmark, our algorithm has the best performance compared to other state-of-the-art real-time object tracking algorithms.