| Thermal infrared target tracking technology is a research hotspot in automatic target recognition(ATR)field.Due to the lack of clear texture and contour features in thermal infrared images,it is difficult for traditional tracking methods to extract effective target features,which is often interfered by similar objects and occlusions in tracking tasks,resulting in tracking errors.To solve the above problems,this paper proposes a deep learning method for thermal infrared target tracking based on representational learning theory.The main work is as follows:(1)Aiming at the problem that the thermal infrared target is interfered by similar objects in tracking process,a hierarchical pre-training feature extraction network with attention mechanism is designed.The fundamental reason that thermal infrared target tracking is susceptible to analogue interference is deeply analyzed.On this basis,the layered pre-training method is used to enhance the feature extraction ability of the network,and the key information of the target is further extracted by combining the attention mechanism.To a large extent,the difficulty of tracking failure caused by analogue interference is avoided.(2)Aiming at the problem that the object is partially occluded in the tracking task,a dynamic template library is proposed.It is not reasonable to use the initial template frame to match the morphologically changed target when the thermal infrared target’s apparent shape changes.In this paper,a target tracking algorithm based on the construction of dynamic template library is proposed to timely update the template when the target is partially blocked and affected by illumination,which effectively avoids the phenomenon of target tracking drift.(3)To verify the proposed algorithm,an ablation experiment was conducted on a public thermal infrared test set.Compared with the baseline algorithm,the success rate and accuracy rate of the target tracking algorithm using the layered pre-training feature extraction network with integrated attention mechanism are increased by 2.8% and 4.7% respectively.The target tracking algorithm based on dynamic template library increases the success rate by 1.2% and the accuracy by 3.8% in the thermal infrared target tracking data set.Experimental results show that the success rate and accuracy of the proposed tracking algorithm are improved effectively. |
