| Infrared cameras have been used more and more widely nowadays,but the research on algorithms for infrared images is still very scarce,especially for infrared target tracking.There are usually two ways to solve the problem of target tracking in infrared image.One is to use hand-craft features,such as HoG.Because there are many assumptions about hand-craft features,the algorithms are rarely used in the real world.The other is to apply the deep learning model for solving RGB images directly to infrared images.It can be imagined that the performance of the model is dramatically reduced compared with visible scenes.According to the two evaluation indexes of target tracking task,accuracy and robustness,two different schemes are studied in this paper.Aiming at the shortcomings of the existing infrared target tracking algorithm based on deep learning,a multi-layer feature similarity model is proposed by analyzing the features of infrared images.For infrared object tracking,different levels of features are good at solving different problems.Low-level features are more robust to target deformation and high-level features are more robust to complex background interference.This paper presents a multi-layer feature similarity model based on Siamese network.The model includes a structure correlation similarity network and a semantic correlation similarity network to calculate the similarity of features at different levels.Structural correlation similarity network can capture the local structure information and help the network distinguish similar thermal infrared targets.Semantic correlation similarity network can enhance the ability of global semantic expression.The model also includes an adaptive integrated network based on relative entropy to fuse structural similarity and semantic similarity.In order to solve the problem that the tracking algorithm based on siamese network can not adapt to the change of target's shape,an adaptive anchor module is integrated into the siamese network based regional proposal network.The anchor frame is generated online by the contour information of the target as a baseline.RPN adjusts the anchor to get more accurate prediction by regressing the error between the anchor and the target.At the same time,in order to make use of the target's motion information,a Bayesian decision module is introduced to adjust the probability of the possible position of the target according to the motion information,which makes the tracking model more robust.In this paper,VOT-TIR2015 and VOT-TIR2017 are used to evaluate the two improved schemes.Experiments show that the proposed schemes can significantlyimprove the performance of the algorithm.The first model proposed in this paper is9.4% higher than its baseline in VOT-TIR 2017,and the second model is 16% higher than its baseline. |
