Research On UAV Visual Object Tracking Method Based On Feature Fusion And Attention Mechanism

Unmanned aerial vehicles(UAVs)have considerable research and application value in scientific research,military,and civil fields.Visual object tracking based on UAVs has received much attention due to its versatility.However,there are challenges such as object deformation,partial occlusion and similar object interference in real environments.Moreover,due to the limited computational resources of UAV onboard processors,most trackers are difficult to balance the tracking accuracy and speed.Therefore,based on the idea of how to improve the matching recognition ability of trackers with real-time speed,this thesis proposes a UAV object tracking algorithm based on feature fusion and attention mechanism.The main contributions of this work can be summarized as follows:(1)A feature matching fusion network combining pixel-wise correlation and dense connection is proposed.To address the problem that the traditional correlation method has difficulty coping with object occlusion when the background is vast,this thesis introduces a pixel-wise correlation to attenuate the background noise,and shows that the pixel-wise correlation has stronger correlation matching ability by feature map visualization.The computational volume of pixel-wise correlation is 6.230 M,which consumes less computational resources.Based on this,a feature matching fusion network combining pixel-wise correlation and dense connection is designed using the dense connection pattern of Dense Net.The network can enhance feature propagation and enrich features.The network complexity is low with the computational volume of79.173 M and the parameters number of 117.12 K.(2)An attention module that fuses spatial and channel dimensions is proposed.Aiming at the problems of blurred features of small targets and similar object interference in the UAV view,this thesis enhances the key features of targets and improves the object characterization by fusing self-attention in spatial dimension and channel attention in channel dimension.The computational volumes of self-attention and channel-attention are 212.686 M and 238.592 K,and parameters numbers are 98.7K and 32.8 K with low computational consumption.In addition,a classification and regression module is designed by borrowing from the lightweight object tracking algorithm,with a computational volume of 3.784 G and a parametric number of 5.598 M.(3)Combining the proposed module,a UAV object tracking algorithm based on feature fusion and attention mechanism is proposed.The precision is 0.697 and the success rate is 0.497 on UAV123@10fps,and the tracking performance is excellent with the precision of 0.723 and success rate of 0.521 on UAV20 L.The computational volume tested on NVIDIA TITAN RTX is 9.597 G and the number of parameters is 9.498 M.The running speed reaches 100.1 fps with low complexity,which can meet the real-time requirements of UAV object tracking.(4)The multi-scene UAV visual object tracking system is built and the proposed tracking algorithm is experimentally verified under real flight scenarios.The results show that the tracking algorithm proposed in this thesis can cope with scenarios such as viewpoint change,object occlusion and similar object interference,and the built tracking system can achieve stable object tracking on UAVs.