Research On Object Tracking Algorithm Based On Siamese Networks In Complex Scenes

As an important research direction in the field of computer vision,object tracking has long been valued and studied by scientists and related scholars.It has broad application prospects in both civil and military fields,and after years of development,many excellent object tracking algorithms have emerged.But in the tracking process,there are many kinds of complex scenes such as occlusion,background clutters,fast motion,out-of-view,deformation and low resolution.These unpredictable factors make it a challenging task to design object tracking algorithms with excellent tracking performance.In this paper,the siamese network Siam RPN object tracking algorithm is selected as the benchmark algorithm to address the above problems,and the main research contents are as follows.1)The current object tracking algorithms based on siamese network in complex scenes such as occlusion,background clutter and low resolution have the problems of poor adaptability and serious drift of the tracking frame.A tracking algorithm combining the attention module and selective kernel module is proposed.It introduces an effective channel attention module and a normalization-based channel attention module.The former attaches importance to the feature information that needs to be extracted,while the latter strengthens the salient features.In addition,the four branches of the region proposal network also incorporate the selective kernel modules,this method can adaptively adjust the receptive field and enhance the global information embedding.Experimenting under the above proposed complex scenarios,the tracking precision and success rate of the proposed algorithm are higher than the benchmark algorithm,which demonstrates the effectiveness of the proposed algorithm.2)There exist some problems in the benchmark algorithm,such as the number of layers of feature extraction network being too shallow to extract more discriminative features,spatial location information and background information being easily ignored,and fixed target templates are not updated,etc.A tracking algorithm combining coordinate information and template update is proposed.Firstly,the improved Mobile Net V3 S network is used as the feature extraction network and multi-layer feature fusion is used to extract richer feature information.Then the coordinate attention module is incorporated to focus on the location and spatial information of the object.Finally,the template update module is introduced,which combines the initial template and the current frame,to adaptively update the template for the next frame.The experimental results show that,compared with the benchmark algorithm,the proposed algorithm significantly improves tracking performance in a variety of complex scenarios such as illumination variation,motion blur and rotating objects,which indicates that the combination of the deep network with coordinate attention and update templates can significantly improve tracking performance.3)Long-term tracking is a challenging branch of object tracking.Aiming at the problem that the target disappears from the field of vision for a long term and is difficult to track when it reappears,a long-term tracking algorithm is proposed,which introduces the aforementioned template update module and combines the target disappearance judgment mechanism with the global search mechanism to determine whether there is a target in the tracking frame and expand the range search for the disappeared target.The long-term tracking dataset and the author’s homemade dataset are selected for the experiment,experimental results show that compared with the benchmark algorithm,the proposed algorithm has better long-term tracking performance.