Deep Learning Based Target Detection And Tracking Methodology Research

In recent years,deep learning has been widely used in remote sensing images to enhance the performance of target detection and tracking due to its powerful feature extraction and representation capabilities.With advantage of capturing ground information,remote sensing images can accurately reflect various real scenes.They are not only widely used in civil applications serving transportation,agriculture and forestry,and environmental monitoring,but also can be used for target reconnaissance,battlefield dynamic analysis,and strike effect evaluation in military applications.However,remote sensing images in real scenes have a variety of situations such as large differences in target size,high background complexity,small size targets,and dense targets,resulting in poor performance of targets in the detection and tracking process,and often wrong and missed detection.In view of above situations,this thesis conducts research based on convolutional neural network target detection and tracking techniques,and conducts experiments under multiple types of remote sensing datasets to verify the effectiveness of targets in multiple scenarios.In this regard,the main research of this thesis is as follows.(1)In the process of target detection,the detection of small targets requires high resolution and sufficient semantic information,while the detection of large targets requires a larger perceptual field.To address this problem,this paper designs a parallel branching network based on scale invariance,and processes the two branching networks separately by using different expansion convolutions,so as to present the perceptual fields suitable for targets of different scales and improve the resolution of small targets while increasing the perceptual fields of large targets.(2)To address the problem of poor characterization of small targets,this paper proposes an enhanced balanced feature fusion approach,which uses the output of the feature pyramid network for fusion processing to obtain global feature information of the image,and then fuses it with the output feature map of the feature pyramid.The model not only enhances the output of the original feature pyramid,but also enables it to incorporate global information.(3)For the nonlinear problem of object motion trajectory,BLSTM motion trajectory prediction model based on recurrent convolutional network is constructed.Based on the original Deep Sort algorithm,BLSTM with regression capability is introduced,which not only effectively obtains the spatio-temporal information of sequence images,but also fuses the contextual information to improve the ability of dealing with the target occlusion problem.The proposed methods in this thesis are tested experimentally under multiple types of remote sensing datasets to verify the effectiveness of the methods and improve the robustness of remote sensing image target detection and tracking to a certain extent.