Siamese Network Visual Tracking with Muti-Layer Feature Fusion and Attention

Object tracking based on Siamese networks has made significant progress in recent years. Convolutional networks can extract multi-layer deep features from images, but some methods only use the last layer feature or use multi-layer features in a simple way, which does not fully exploit the advantages of features. In this study, a new multi-layer feature fusion module is proposed by using different dilation rates dilated convolutions to expand the receptive field and an attention mechanism is introduced to further enhance feature representation. Moreover, we optimized the selection strategy of positive and negative samples. Extensive experimental evaluations were performed on OTB100 and UAV123 benchmarks to validate the effectiveness of the proposed tracker. The results indicate that our method not only achieves substantial performance gains but also satisfies realtime tracking requirements. The success is increased by 1 % on UAV123 benchmark.