Research On Vision Detection And Tracking Technology Of UAV Aerial Vehicles

Currently,UAV aerial photography has become an important means for people to obtain information.The research on the visual detection and tracking technology of aerial targets is of great significance to the fields of military investigation and public safety.However,different from the ground perspective,the scene under the drone’s aerial perspective usually has the characteristics of complex background and small target.In addition,problems such as camera movement and occlusion of the target are more common,which brings a huge challenge to the visual detection and tracking algorithm.This paper studies the YOLO detection algorithm and the regularization correlation filter tracking algorithm.Taking the drone aerial vehicle as the main experimental object,the overall framework and process of the visual detection and tracking algorithm in this paper are constructed,and then the visual detection and visual tracking algorithm is designed and improved.The main work and contributions of this paper are as follows:(1)Designed the YOLO visual detection algorithm that integrates the attention mechanism.Aiming at the problem of YOLOv4-tiny’s fast speed but low accuracy,the network is improved by fusing the attention mechanism of its feature extraction network.Model training,ablation experiments and comprehensive tests were carried out on the UAV aerial photography data set Visdrone.The experimental results show that the speed of the improved algorithm has dropped by 3.8%,but the mean Average Precision has increased by12%.In addition,experiments on self-built data sets show that the improved algorithm has better generalization ability.(2)Propose a regularized correlation filtering visual tracking algorithm fused with scene perception memory.The global information of the relevant filter response graph is used to characterize the tracking confidence,combined with the local changes of the response graph,the historical view changes where the target is located,and the historical filter information,a scene-aware memory mechanism is designed,so that the tracking algorithm can perceive rapid target movement,camera movement,occlusion,etc.The tracking anomaly brought by it,and rapid re-detection at the same time,prevents model pollution and tracking drift.Based on the scene-aware memory mechanism,in response to the high accuracy of the regularizationrelated filtering algorithm but the poor real-time performance caused by the complex model,three optimizations are carried out in three aspects: adaptive time regularization,adaptive model sparse update,and adaptive feature sampling.Simplifies the model and speeds up the solution.(3)Based on the unified evaluation standards,extensive experiments of visual tracking algorithms have been carried out on UAV123@10fps,DTB70 and UAVDT UAV aerial photography data sets.The overall performance evaluation shows that the tracking algorithm in this paper shows the best tracking performance on the aerial vehicle data set UAVDT.The evaluation based on the scene attributes shows that the tracking algorithm in this paper has better performance in occlusion and camera motion scenes.The ablation experiment shows that the tracking algorithm designed in this article With a 1.6% loss of accuracy and a 1.7%reduction in success rate,an acceleration of nearly 1.7 times is obtained;the robust tracking experiment of aerial vehicles shows that the tracking of this text is integrated with scenesensing memory in the scenes of target occlusion,out-of-view,fast motion,and camera motion.The algorithm performs better.In short,the visual tracking algorithm designed in this paper has achieved real-time and robustness improvement with a certain degree of accuracy and success rate reduction.