Scale and appearance variation enhanced siamese network for thermal infrared target tracking

Thermal infrared target tracking has recently attracted much attention. Although various methods have been proposed, they are usually sensitive to the target appearance and scale variation which will lead to tracking deviation. Therefore, in this paper, we propose a novel scale and appearance variation enhanced siamese network to solve the above problem. First, we introduce the dilated convolution module into siamese network to enlarge the receptive field of network while maintaining the resolution of feature maps. Hence, the model adaptability to the target with different scales could be enhanced. In addition, we devise a target template library update strategy based on the tracking results of historical frames to reduce the dependence of the network on the initially given exemplar. In this way, the appearance change of targets could be better described so as to overcome the tracking deviation problem over time. Qualitative and quantitative results on PTB-TIR, LSOTB-TIR and VOT-TIR2015 datasets demonstrate that our proposed model is able to achieve superior thermal infrared target tracking results. Especially, when the target appearance and scale have significant variations, our model shows outstanding performance than other state-of-the-art methods.