Object Tracking Research Based On Feature Fusion Fullyconvolutional Siamese Networks

As one of the hot research directions of computer vision,target tracking plays an irreplaceable role in the fields of automatic driving and intelligent monitoring.Due to the wide application fields and the complex application scenarios of target tracking technology,scholars from different fields have proposed a variety of target tracking algorithms.The Siam FC(full convolutional Siamese networks for object tracking)has been concerned by many researchers since it was proposed because of its excellent tracking speed.However,the tracking success rate of Siam FC algorithm is not enough,and the feature discrimination ability of this algorithm is insufficient.In order to improve the ability to discriminate and locate targets,the feature extractor of this algorithm will be improved as follows:Since the ability of distinguishing features extracted from the backbone network of Siam FC algorithm is insufficient,this paper improves the basic network of this algorithm and proposes a new object tracking algorithm SF?Siam(Shuffle Net?based Siamese).In addition,we design a cropped no-padding residual unit to replace the original unit of Shuffle Net to eliminate the bad influence of deep network padding on Siamese network target tracking algorithm.And we readjusted the structure of Shuffle Net according to the total stride of the network and the size of the output features.The test results on OTB(object tracking benchmark)and VOT(visual object tracking)data sets show that SF?Siam algorithm can effectively improve the ability of feature discrimination.In order to further improve the tracking success rate and tracking accuracy of SF?Siam algorithm,we paper proposes a three branch Siamese network object tracking algorithm CSAD?Siam.It retains the advantages of SF?Siam algorithm,and adds channel and spatial attention module as well as deformation attention module to improve the expression ability of features.In the channel and spatial attention module,channel attention mechanism and spatial attention mechanism are added to redistribute the weights of different channels and the space regions of the feature map,so that more and more attention is paid to the information which is beneficial to target tracking.In the deformable attention module,the output features of different convolution layers are fused to obtain more abundant feature information.Meanwhile,the deformable attention mechanism is added to enhance the ability of the model to deal with the geometric transformation of the target.Experiments show that the CSAD?Siam algorithm can effectively integrate the advantages of each branch module and improve the tracking accuracy.