Research On Visual Tracking Of Multiple Ground Targets For Multi-UAVs

With the rapid development and improvement of drone technology,single-drone tracking of ground targets has been difficult to meet increasingly complex execution tasks,and multi-drone collaborative work has gradually become a main presentation method of drones in practical ap plications.For example,there are drone groups for civilian use to implement traffic supervision,and many drones for military use to implement coordinated tracking and strike against ground targets.At present,with the widespread application of computer vision technology in various fields,combining vision technology with drones to achieve target recognition and tracking has great application value in many fields,which is also a major research direction for future drone platforms.This dissertation is mainly aimed at solving multiple UAVs using visual tracking technology to complete the tracking tasks of multiple ground targets in complex scenarios.Multi-aircraft multi-target tracking is the distribution and tracking of multiple ground targets through information interaction between drones.It is mainly divided into two parts: coordinated interaction of drone groups and tracking and positioning of a single drone platform.In this paper,computer vision technology is used to complete target positioning a nd tracking on a single drone platform.Currently,most commonly used visual tracking algorithms use traditional image methods or deep networks to extract features from targets,and then use classifiers to distinguish between targets and backgrounds.Locat ion tracking.Multi-UAV multi-target allocation can be divided into centralized allocation,distributed allocation,mixed allocation and other methods according to different scenarios and applications.Heuristic algorithms are mainly used to implement the optimal allocation scheme.This dissertation improves the traditional correlation filtering target tracking algorithm,adopts multi-feature weighted fusion,feature extraction method of weight factor self-learning to complete the target feature extraction,and uses a class classifier to classify the target and background in the image.In order to reduce the signal-to-noise ratio of feature extraction,a spatial confidence mask is introduced to improve the accuracy of target positioning from the pixel level.In order to take into account the tracking speed and accuracy,this paper proposes an adaptive scale estimation based on feature layer for multi-position search.In terms of multi-aircraft multi-target tracking,UAVs are susceptible to the limitations of its on-board resources and natural factors.Therefore,multi-target allocation and tracking is a target optimization problem subject to multiple conditions.In this paper,the centralized target pre-allocation based on genetic algorithm and the target redistribution based on extended contract network algorithm are used.During the tracking process,according to the change of the environment type and the change trend of the drone's own revenue,multiple drones alternately change the tracking target to achieve a stable improvement in the overall tracking effect of the drone group and maximize the value of information obtained.In this dissertation,the target tracking algorithm is verified on the UAV123 data set.The tracking accuracy and tracking success rate are used as evaluation indicators and compared with the traditional algorithms,and a certain improvement effect is obtained.The algorithm is simulated in terms of UAV target allocation,and the cost function curve of the allocation scheme and the change of the overall revenue of the UAV group are analyzed.The effectiveness and rationality of the algorithm in pre-allocation and redistribution are verified.