Searching And Technology Of Remote Sensing Target Tracking In Unmanned Aerial Vehicle Videos Based On Depth Perception Modeling

With the gradual development of remote sensing target tracking technology based on Unmanned Aerial Vehicle videos,it has become one of the main research directions in the field of visual tracking and has been widely used in military and civilian fields.However,due to the characteristics of the drone platform,the video sequences have difficult problems such as low resolution,complex and changeable background,and many rotation changes of highly dynamic targets in the field of view,typical target tracking methods for natural scenes cannot be used directly,and their robustness and adaptability are difficult to guarantee.To address these problems,this paper proposes an algorithm for remote sensing target tracking in Unmanned Aerial Vehicle videos based on depth perception modeling,technologies such as data enhancement based on structured information deletion,feature enhancement extraction based on multi-frequency feature representation,contextual information acquisition based on Criss-Cross enhanced attention model,and rotation resistance enhancement based on variable angle adaptation are carried out progressively.The specific research work is summarized as follows:First,in view of the limited training sample data and most of the existing typical augmentation methods don’t consider the characteristics of the scene,this paper adopts the data augmentation method based on structured information deletion,which equivalent to adding a regular term to the network to avoid over-fitting in the network.It also simulates the occlusion during the stage of training.This method increases the amount of data and makes the trained model more generalized,which provides data guarantee for the tracking network.Second,for the problem that the target is difficult to be captured during the stage of tracking caused by the low-resolution feature of Unmanned Aerial Vehicle videos.this paper adopts the feature enhancement extraction method based on multi-frequency feature representation,presents a new multi-frequency feature representation method and introduce the octave convolution into the Alex Net architecture to adapt to the new feature representation,thereby effectively improving the feature expression capabilities for highly dynamic targets while reducing memory consumption and computational cost;For the problems of complex background and many similar interfering targets in Unmanned Aerial Vehicle videos,this paper adopts the contextual information acquisition method based on Criss-Cross attention model,and realizes the enhancement of the connection between modules,effectively adding additional visual information in addition to its own feature information.These two methods effectively improve the robustness of the tracking network by means of internal enhancement of the network.Thirdly,in order to solve the problem that the highly dynamic targets frequently rotate during the stage of tracking,which may cause the tracking result drift,this paper constructs a new detection-tracking framework and adopts a rotation resistance enhancement method based on variable angle adaptation.The two-stage operation of information acquisition and angle consistency update enables the tracking network to realize the perception and adaptation of the target angle change.This method effectively improve the adaptability of the tracking network by strengthening the support of the external architecture.Finally,through the above technical support,this paper forms the full-link remote sensing target tracking algorithm in Unmanned Aerial Vehicle videos based on depth perception modeling.In this paper,with the support of Unmanned Aerial Vehicle remote sensing datasets,qualitative and quantitative analysis is carried out,and multi-level and multi-angle verification means are used to verify the effectiveness of the method proposed in this paper.