Research On Object Tracking Based On Siamese Network

With the emergence of massive video data and the popularity of artificial intelligence,visual object tracking is highly promising in military and civilian applications and it has received more and more attention in academia.However,the objects' priori information is less,and the target itself as well as the tracking scene may have some unpredictable changes in subsequent frames,so object tracking algorithms still have a lot of room for improvement in terms of robustness and accuracy.In recent years,deep learning driven by big data has extracted better semantic feature representation than traditional manual features,and its performance in computer vision tasks has surpassed traditional algorithms.In the field of object tracking,many researchers have applied deep learning into tracking algorithms to improve the tracking results.Among them,SiamFC(FullyConvolutional Siamese Networks)is banlanced between robustness and real-time performance which utilizes siamese network to extract feature representation and transform the tracking problem into a problem of solving the similarity between the known target and search area.However,the tracking robustness of tracking is reduced in the case of sudden changes in illumination intensity,similar background interference,and obvious appearance changes of the target.In view of the problems existing in SiamFC,this paper draws on the tracking framework of siamese network,and proposes two algorithms to respectively improve tracking process and adaptively adjust network.The improvement in the tracking process is mainly divided into three parts:(1)Introducing data augumentation in the training data,so that the network model learns the feature representation that is more suitable for illumination changes.(2)The tracking confidence is defined.When the confidence is high,the update feature template is selected,so that the tracker can obtain the timing information of the target and the background.(3)In order to better distinguish the target and similar background interference,a strategy of fusing different layers of convolution features is used.On the other hand,since the convolutional neural network of SiamFC is trained based on the object detection dataset,and the extracted features are too general,so this paper respectively proposes a channelbased and spatial-domain-based network regulating mechanism to extract more distinguishing features.The two mechanisms adjust the feature representation of the network extraction by adjusting the weights of the channel and the spatial position when tracking different targets,so that SiamFC can better distinguish the target from the background.The test results on the tracking dataset show that the proposed algorithm in this paper can effectively improve the success rate of tracking without sacrificing much tracking speed,and also provide an improved theoretical reference to solve object tracking's problem in similar background interference and significant changes in appearance.