Target Association Technology In UAV Remote Sensing Images

As a key technology of UAV scene perception and understanding,target association in UAV remote sensing image focuses on solving the corresponding relationship of target data in remote sensing image obtained at different time and different locations.It has a broad prospect in intelligent security and military reconnaissance fields.Especially,the technology of vehicle-target association is of great significance to suspect vehicle search,cross-border vehicle tracking and traffic behavior analysis.From the perspective of UAV,vehicle target association mainly faces the challenge of high similarity between classes.How to extract effective discriminative features is the core of this paper.With the development of deep learning theory,it is a trend to use convolutional neural network with strong representation ability to extract vehicle target features.However,the current target association algorithm based on deep network mainly has the following limitations: 1)it is highly dependent on auxiliary annotation information,2)it does not adapt to UAV application scenarios,and 3)it does not extend the association algorithm for vehicle target images into complex scenes.Therefore,aiming at the above problems,this paper focuses on the extraction method of target discriminative features based on deep learning,improves the association performance of target association model from the perspective of UAV,and extends it to the process of vehicle target association in complex scenes.Firstly,in order to reduce the dependence of target association algorithm on auxiliary annotation information,a vehicle target association algorithm based on global feature is proposed.Using color prior information with strong generalization and advanced semantic information of depth feature map of neural network,an attention mask was generated based on PCA algorithm to emphasize target region and suppress background noise to guide the association network to extract discriminant features.A weighted triplet loss function is proposed for UAV’s special imaging perspective,which alleviates the problem of large similarity between vehicle target classes to a certain extent.At the same time,involution operator is introduced to further improve the vehicle target association performance on the basis of reducing the number of associated network parameters.Secondly,in order to improve the granularity of target representation from the perspective of UAV,a vehicle target association algorithm based on local features is proposed.A self-aligned local feature extraction algorithm was proposed based on the analysis of vehicle image features from UAV’s special imaging perspective,which aligns input images without the need for perspective annotations,and integrated global feature and local feature to improve the representation ability.Then,the relationship network is used to deduce the relationship between local features to avoid confusing different vehicle targets with similar local attributes and improve the performance of target correlation.A large number of experiments on UAV-Ve ID show that the proposed correlation algorithm can be adapted to UAV remote sensing application scenarios.Finally,this paper integrates the above association algorithm into the target association flow in complex scenes,and designs corresponding algorithm flow for single target association and multi-target association respectively.For the former,the similarity threshold is determined according to the experimental analysis.For the latter,a new R similarity measurement method is proposed and modeled as a binary graph optimization problem.The performance of target association algorithm is tested on UAV-ROD and outdoor real UAV remote sensing images.