Siamese Neural Network Models For Thermal Infrared Object Tracking

Thermal infrared object tracking is a sub-task of visual object tracking.Thermal infrared obect tracking can be applied to video surveillance in the harsh weather conditions,aerospace industry,military industry and other fields.The main difference between visual object tracking and thermal infrared object tracking is the difference in input images.Due to imaging techniques,thermal infrared images contain many white noises,blurred images,and insufficient texture information.Due to these characteristics of the thermal infrared image,there is a large performance loss when directly applying the tracker to the thermal infrared object tracking,which is caused by the lack of the underlying feature extraction method of the tracker.The significant discriminative information about the object that can be observed in the thermal infrared image,is the edge structure of the target.The most commonly used HOG features in the visual target tracking do not make full use of the amplitude information of the gradient,so the edge structure of the target cannot be accurately described.Inferior performance in infrared target tracking.In recent years,the deep features commonly used in the field of computer vision are features extracted by VGG-Net trained on visible light image data sets.Since the information contained in visible light images and thermal infrared images is different,when applyed this feature extraction method to thermal infrared object tracking has performance loss.Aiming at these problems,this paper proposes to introduce a feature extraction method based on local adaptive kernel regression into the thermal infrared object tracking,and improve the feature extraction part of the benchmark tracker ECO.This paper refers to the improved ECO tracker as ECO._LSK.This paper compares the shortcomings of traditional feature extraction methods,and analyzes the reasons why this feature extraction method performs well in thermal infrared object tracking.In this paper,a comparative experiment was conducted on the thermal infrared target tracking evaluation data set VOT-TIR2017.The ECO_LSK tracker proposed in this paper surpassed the HOG feature ECO tracker 25% in the EAO tracker,and exceeded all the ECO tracker with traditional feature extraction methods.The feature extraction method proposed in this paper cannot meet the real-time requirements of the tracker because of the long calculation time.Aiming at this problem,this paper proposes a siamese neural network model combined with correlation filtering algorithm as a tracker for thermal infrared target tracking.Aiming at the problem that the high-level semantic features extracted by deep network are easy to be confused,this paper proposes to adopt a shallower and lighter network design,and transplants the underlying feature extraction network to the tracker based on the correlation filtering algorithm,then proposes the tracker T-CFNet.Aiming at the problem of insufficient training data set in thermal infrared target tracking,this paper uses a model based on generative adversarial networks to generate the required training data from the general visual object tracking data set,and discusses the training data scale impact on tracker performance.In the thermal infrared object tracking evaluation data set VOT-TIR2017,a comparative experiment was carried out,and a variety of classic trackers and ECO_LSK trackers were compared,which proved that the T-CFNet tracker achieved ideal accuracy,robustness,and speed.It overspeeds the ECO_LSK tracker by 9.8 times and can be directly applied to industrial production.