Research On Target Tracking Method Of UAV Aerial Photography Based On Siamese Networks

As one of the bottom key technologies of the computer vision system,visual object tracking has no special restrictions and requirements for the category,shape and motion model of the object to be tracked,so it has a wide application prospect.In recent years,through the introduction of deep learning technology,the performance of tracking algorithms has been greatly improved.Some newly proposed end-to-end,unsupervised deep tracking models solve the problem of poor real-time performance of deep learning-based tracking methods to a certain extent.The fast development of UAV technology and the rapid improvement of computing processing capability provide a wide range of application scenarios for UAV target tracking technology.However,in the UAV tracking task,the instability of the UAV platform has caused many uncertainties.Challenges such as background clutter,similar target interference,and frequent occlusion also increase the difficulty of target tracking,which lead to a lot of performance degradation of the recent state-of-art algorithms.In order to promote the application of object tracking method based on deep learning in unmanned combat,intelligent combat and other occasions,robust object tracking algorithms with anti-occlusion and good accuracy were studied.The main work of this dissertation is as follows:(1)The basic modules of Siamese tracking network have been improved.Aiming at the lightweight Siamese network,a cross-layer connection from the feature layer to the correlation layer was added to improve the mobility of information,which made the model has better training effect.The modified Res Net-50 was used as the backbone network to improve the performance of the model in dealing with the target deformation and suppressing background interference.The network was trained by using features of different depths,and it was found that the model with the best performance was that built with features of the p3 layer.In order to fully excavate the potential of deep features,two different fusion methods of correlation features were proposed.The experimental results showed that the coarse-to-fine cascade method is more beneficial for performance improvement than the weighted fusion method.(2)Aiming at the shortcoming of Siamese tracking network’s poor ability to distinguish similar targets,a Siamese Network with Adaptive Template Update(Siam ATU)based on correlation filter was proposed.The research on the model updating method showed that the template construction with shallow features and correlation filter modulation has better model updating effect.By embedding correlation filter into deep network as a layer,end-to-end network training was realized;an adaptive confidence evaluation mechanism was adopted to avoid the pollution of sample pool in case of occlusion.Siam ATU combines the shallow correlation filter response and deep semantic classification,so it has more advantages in aspect of anti-interference.(3)In order to further improve the ability to distinguish difficult samples,a Two-Stage Siamese Network(TSSN)suitable for UAV platform was proposed.Referring to the design ideas of the two-stage network in detection field,through adopting Ro I pooling layers,the features of candidate targets obtained by the region proposal network were extracted and input to the verification network for more refined classification and regression.In the training phase,a joined training method combining the positive and negative sample pairs was used to improve the model’s ability to distinguish difficult samples.In the UAV tracking scene,the target size is generally small,and it is difficult to distinguish the target from distractors only depending the apparent features.Through a multi-trajectory tracking mechanism,the robustness of the model was improved by combining the motion information and depth feature information.The experimental results showed that compared with the one-stage Siamese network,the classification ability and regression accuracy of TSSN were significantly improved under the premise of increasing a small amount of calculation.(4)Aiming at the recapture problem after the target is lost,a long-time tracking framework based on a global proposal network was put forward.The global proposal network adopts a feature pyramid structure design,and uses a target aware module to encode the interaction between the instance feature and the search frame,so as to improve the detection ability of small-scale targets;compared with long-time tracking methods adopting a sliding window operation,it has higher calculation efficiency and detection ability.Aiming at the problem of camera shaking caused by the instability of the UAV platform,the pyramid L-K optical flow was used to calculate the projection transformation matrix of the background,and the projection matrix was used to modify the Kalman filter,so that the multi-trajectory tracking mechanism is still applicable.The experimental results showed that the performance of the proposed method on UAV datasets is better than other comparison algorithms.The method is universal,which provides a good solution for search and location of small-scale targets in the tracking scene with a large field of view.