Research On UAV Object Tracking Methods Based On Siamese Network

Object tracking for UAV is an important part of the Anti-UAV task,and at the same time,object tracking algorithms based on fully convolutional Siamese networks have been developed rapidly due to the balance of accuracy and real-time performance,so it becomes feasible to apply Siamese networks object tracking algorithms to Anti-UAV task.However,there are still many difficulties in the actual object tracking for UAV.Firstly,there is a lack of Anti-UAV data and a single type of scene.Secondly,the complex background of UAV low-altitude flight and the small pixel share of the UAV object in the video image makes it more difficult to track successfully.Finally,the conventional tracking algorithm will show tracking drift and poor tracking performance for UAV in long-time tracking situations such as occlusion or flying out of the framing range.To address the above problems,the following solutions are proposed in this paper:(1)In response to the problem of insufficient amount of Anti-UAV data and a single type of test scenario,the homemade Anti-UAV dataset HEU Anti-UAV was filmed,edited,analyzed and labeled by actual,and further divided into 9 tracking scenarios,which can simulate the possible problems among actual Anti-UAV missions and effectively compare the Anti-UAV tracker ability.(2)For the Anti-UAV scenario,the tracking algorithm does not perform well for complex background and small object tracking.A Siamese neural network object tracking algorithm with the introduction of an attention mechanism and feature rearrangement called Siam AU is proposed,which is based on Siam RPN++ and combined with an attention mechanism and a feature rearrangement technique.First,the ECA-Net attention module is integrated into the backbone network to enhance the representation ability of the convolution features in the complex background.Then,the channel number of the last three convolution features is rearranged in order to make full use of the low-level features that are conducive for small object tracking.The rearranged feathers are further fused to obtain the improved feature map.Finally,extensive experiments on two public datasets,validate that the proposed Siam AU achieves better UAV tracking performance and outperforms previous methods,especially in small objects and complex background scenarios.(3)For situations such as occlusion or flying out of the framing range of UAV objects that exist in long-time tracking situations,conventional tracking algorithms for UAV objects perform poorly and are prone to tracking drift.Appropriate use of detection algorithms can detect UAV objects,improve the tracking success rate,and provide an effective solution to the above problems.Based on this,two different ways of combining detection algorithms and AntiUAV tracking algorithms for UAV objects are proposed,and the test results on two Anti-UAV datasets show significant improvement compared to conventional tracking algorithms.