| Target tracking is a hot research and application in the field of machine vision.Domestic and foreign scholars are committed to improving the accuracy and robustness of tracking algorithms.Nowadays,with the wide application of UAV in production and life,it has become a development trend to realize the application of computer vision on UAV platform.In this paper,a pseudo siamese network framework is designed,which contains a target tracking branch and a template library branch,which performs target tracking task and template storage update task respectively.Target tracking branch is an improved algorithm based on kernel density estimation to complete the target tracking task.The algorithm of this branch can solve the problems of scale change and background interference in the process of UAV target tracking.At the same time,the target model obtained by tracking each frame is input into the branch of template library for processing.The template library branch completes the selective storage task of the target feature model in the tracking process.The template in the template library was scored,and the appropriate template input tracking branch was selected as the follow-up tracking target feature model to solve the problem that the feature fitting ability of the target model in the initial frame gradually weakened.The work of this paper can be summarized as the following aspects:In the target tracking branch,an improved algorithm based on kernel density estimation is used to track the target.For uav target dimension changing in the process of target tracking problem,the candidate targets using variable bandwidth characteristics of kernel density estimation to express,through to the target features and candidate model similarity calculation of scale adaptive adjustment.Secondly,two regularization terms were introduced into the estimation scale,and the rationality of the estimation scale was verified by backward consistency detection,and the scale was determined according to the verification results.Aiming at the uncertainty of the flying height of UAV,a large amount of background noise is introduced into the target tracking box,and the target is greatly affected by the background noise,and the scale ambiguity of the self-similar object is high.In order to reduce the influence of background information on the tracking effect,a similarity measurement based on the weight of background pixels was proposed,and background pixel histogram was introduced into the calculation of similarity measurement,which enhanced the tracking accuracy of the algorithm under complex background.In the template library branch,the characteristic model of the target tracked by the target tracking branch is selectively stored to establish the template library.Kernel density estimation method uses the feature model of the initial frame to run through the whole tracking task.In the process of tracking,the fitting ability of the feature model of the initial frame is gradually weakened,which easily leads to problems such as difficult target matching.To solve this problem,this paper proposes a template library branching strategy of pseudo-twin network framework,which selectively stores the target feature models obtained in each frame of the tracking branch into the template library,and selects the appropriate template input tracking branch as the target model of subsequent frames through the established scoring mechanism.Practical testing of UAV target tracking.According to the application scenario of tracking algorithm,power consumption,UAV volume,embedded hardware and other aspects of consideration.The hardware and software basic platform of UAV is built.Then the algorithm is transplanted to the UAV platform and the actual test of the UAV target tracking algorithm is carried out.The test results show that the system can carry out the scale adaptive tracking of the target in the actual flight. |
