Research On Visual Object Tracking Algorithm Based On Siamese Neural Network

Visual object tracking,as one of the research hotspots of computer vision technology,plays an important role in many fields,such as intelligent monitoring and human-computer interaction.There are many applications in object tracking,and the application scenarios are also very complicated.Therefore,tracking failures often occur for various reasons.Researchers in different fields have proposed many different tracking algorithms.The SiamFC(Siamese Fully-Convolutional)object tracking algorithm based on siamese neural network has received lots of attention since it was raised.However,the tracking accuracy of this algorithm is not enough,in order to improve the ability to discriminate and locate targets,some improvements to this algorithm are proposed in different aspects.In order to solve the shortcomings of feature discriminating ability and the inability to accurately distinguish background interference in SiamFC,the basic network structure is improved.In this thesis,a no-padding inside cropped residual unit is explored and a number of different inside cropped residual network structures are constructed from three aspects: stride size,maximum receptive field and the output feature size of the convolutional neural network.Based on the above network structures for object tracking,a SiamICD(Siamese Inside Cropped Darknet)tracking algorithm is proposed.The SiamICD algorithm is tested on the OTB(Object Tracking Benchmark)dataset.The experimental results show that the proposed algorithm improves the discriminating ability of features.In order to further enhance the performance of SiamICD,the SiamICD-AFU tracking algorithm is proposed to improve the similarity decision and update method of SiamICD.In terms of similarity decision-making,the channel wise cross-correlation method is presented.Meanwhile,the feature spatial attention mechanism and the feature channel attention mechanism are introduced.As a result,more and more attention has been paid to feature space positions and feature channels which are beneficial to object tracking.In addition,the confidence maps obtained by low-level features and high-level features are weighted and combined to get a more accurate position of object.In terms of update method,a linear model update method based on the confidence of the confidence map is introduced.The tracking accuracy and success rate in the OTB-2015 dataset of SiamICD-AFU tracking algorithm proposed in this thesis is 0.882 and 0.656.Compared with the original SiamFC algorithm's 0.771 and 0.582,the performance of the original one is effectively improved by the proposed method.