Research And Application Of UAV Object Tracking Based On Siamese Network

Applying object tracking technology to UAV platforms is a key problem in computer vision,which has a wide range of applications in battlefield reconnaissance,border patrol,intelligent transportation,power grid detection,and disaster detection,etc.With the efforts of scholars at home and abroad,the object tracking theoretical framework is becoming more and more mature,but it is still difficult to apply it to the UAV platform.In addition,due to the influence of many uncertain factors in the tracking process,designing a fast and robust UAV object tracking algorithm is extremely challenging.Therefore,researching UAV object tracking not only has important theoretical significance,but also has wide application prospects and high practical value.Since the emergence of the SiamFC algorithm in 2016,the object tracking algorithm based on the siamese network has achieved excellent results in various competitions,and the researchers have gradually shifted the direction of the object tracking algorithm from related filtering to the siamese network.Full convolution siamese network tracking algorithm(SiamFC)achieves real-time speed and meets the basic conditions for target tracking applications on UAV platforms.However,SiamFC’s feature extraction network is shallow AlexNet and lacks an online update process,resulting in a high probability of tracking failure when the algorithm changes dramatically in appearance of the target.In addition,there is not much energy on the UAV platform to support a large number of operations.In view of the above problems,this thesis improves the SiamFC algorithm in the framework of twofold networks.The main contents of algorithm improvement are as follows:(1)Changing the feature extraction network in SiamFC and replacing the original AlexNet with DenseNet.DenseNet has simple structure and less computation,which can save the energy of UAV,and a common attention module is added to the target image branch of the siamese network to capture the local features of the target.In this way,the object tracking speed is guaranteed and the accuracy of network tracking can be improved;(2)Proposing a siamese network model consisting of the appearance network and the semantic network to enhance the discriminative ability of the network in the tracking process.The appearance network is undertaken by the dense siamese network in(1).The feature network in the semantic network uses VGG-Net.The channel attention module and the spatial attention module are added to the target branch to emphasize the local features of the target,and the feature fusion module is added to improve the ability of network representation.Finally,in order to verify the effectiveness of the algorithm,a comparative experiment was performed on the UAV-123 dataset.Experimental results show that the improved algorithm in this thesis has better robustness and higher accuracy under the premise of ensuring the tracking speed.There are 16 figures,8 tables and 71 references in this thesis.