Visual Tracking Algorithm Based On Attention Mechanism

With the construction and development of smart cities,computer vision-based target tracking technology has become increasingly prominent and has been rapidly developed.It has been widely used in video automation monitoring,traffic monitoring,virtual reality and other fields,and has become a field of computer vision.Important research topics and research hotspots.In recent years,the research of target tracking based on computer vision has made great progress.However,in the process of target tracking,background interference,occlusion,changes in target attitude,and changes in environmental lighting all affect the results of tracking.Tracking is still a challenging task.Recently,deep learning has become an indispensable model for target tracking algorithms.This article uses MDNet and RT-MDNet algorithms as the research basis.Based on the qualitative and quantitative analysis of the algorithms,the tracking effect of RT-MDNet algorithm is aimed at under complex background conditions.For problems such as poor,this article conducts in-depth research on the target tracking algorithm based on the attention mechanism and has achieved the following innovative results.(1)Propose a target tracking method combined with attention mechanism.This method introduces an attention mechanism on the basis of the RT-MDNet algorithm,constrains the local variance of the features of the input image after convolution,adjusts the CNN backbone network model in the algorithm,and proposes a target tracking combined with the attention mechanism The algorithm strengthens the effectiveness of the feature extraction results,improves RT-MDNet's ability to discriminate targets,and improves RT-MDNet's poor tracking effect under complex backgrounds.(2)A target tracking algorithm based on morphological feature augmentation is proposed.The state of the target will change during the moving process,causing the feature extracted by the RT-MDNet algorithm to change,thereby reducing the tracking accuracy and even causing the tracking failure.In response to this problem,the algorithm proposed in this paper starts from learning the robustness of the target shape,Use the target features in the fully connected layer for learning,and learn the similar features between them,so that different targets can learn similar features in different states,so as to extract the robustness characteristics of the same target in different states.(3)The algorithm in this paper has been tested on target tracking on OTB2013,OTB2015,UAV123 and Temple Color128 data sets,and compared with multiple target tracking algorithms.The experimental results show that the proposed algorithm can better solve the complex background problem in target tracking,and has certain advantages compared with other algorithms.