Research And Application Of Single Target Tracking Algorithm Based On Siamese Network

From traditional algorithms to correlation filters,and then to the algorithm based on deep siamese network in recent years,while tracking accuracy and robustness have been continuously improved,high real-time frame rates can also be maintained.However,deep siamese networks cannot accurately calculate target mask,and lack the ability to filter and strengthen key features,as a result,they have deficiencies in dealing with confusing background interference,occlusion and accurate positioning.The research purpose of this paper is to improve the tracking accuracy and robustness in complex scenarios by introducing an attention mechanism to siamese networks while maintaining real-time frame rate.Based on in-depth analysis of existing deep siamese structures,this paper proposes to fuse attention mechanism into base network and tuning network of Siam Mask to achieve the goal of gradually optimizing network parameters.Then,the improved tuning network is used to test its dual-mode tracking performance in visual and infrared scenes.The main work of this paper is summarized as follows:(1)Aiming at the problem that siamese networks have insufficient ability to deal with confusing background interference and occlusion,this paper proposes to add an attention module to feature extraction structure in Siam Mask.To focus on the key features of target during feature extraction of base network,the attention mechanism module is integrated in its shared siamese structure.Both channel and space constraints help network to strengthen the key target features,suppress the effects of confusing background and occlusion,and provide accurate candidate response information for subsequent foreground classification,bounding box regression and segmentation mask tasks.Test analysis on three VOT benchmark datasets shows that the improved base network has better robustness and EAO.(2)Aiming at the problem of error accumulation caused by inaccurate target positioning in siamese networks,it is proposed to continue to add the attention module to tuning network of Siam Mask.To select the features that are helpful to accurately generate target mask from features which combining low-level spatial details and high-level semantics,this paper adds attention constraints in multiple tuning units in tuning network,thus reducing the possibility of each pixel being wrongly judged after step-by-step upsampling.Both qualitative and quantitative results on VOT 2018 dataset show that the tuning network with attention mechanism improves all of accuracy,robustness and EAO.(3)To realize universal dual-mode target tracking algorithm,improved tuning network is applied to the infrared tracking scene.In this paper,firstly,the infrared dataset taken from multiple angles is made,and the rotated bounding box labels and evaluation criteria are given;Then,the improved tuning network and several typical target tracking algorithms are tested on this dataset.Quantitative and qualitative analysis of results shows that the improved network is feasible for infrared target tracking.