Research On Target Positioning And Tracking Based On Kernel Correlation Filterin

Aiming at the problems of complex solution and difficult to ensure accuracy in the process of visual measurement,and the poor robustness of traditional target tracking algorithms in the process of long-term tracking,this paper studies the target spatial positioning based on multi vision and the target tracking algorithm based on kernelized correlation filter,Combining the two parts of work,a real-time target space positioning and tracking test system based on multi vision is built.The main work of this paper is as follows:(1)First,the camera imaging model and basic theoretical knowledge are described,and the common methods of spatial coordinate calculation are introduced.After comprehensively considering the sources of errors in the actual measurement process and the complexity of the traditional multi-vision measurement method,"Perpendicular Foot Method" is proposed to realize the spatial location of the target,which is simpler and more accurate than others.(2)Based on the idea of feature fusion,a multi-feature weight adaptive position filter training method is proposed,which realizes dynamic adjustment of feature weights through iterative training;Designed a multi-kernelized correlation filtering mechanism,using three filters to jointly estimate the target position,so that the algorithm can accurately distinguish the target and background information during the tracking process,and improve the algorithm’s adaptability to target transformation;In view of the fixed scale problem of KCF algorithm,a scale filter is independently trained to achieve the target’s scale adaptation,and the feature pyramid is used to accelerate the feature extraction during the training process.(3)In order to solve the problem of serious occlusion and target loss in the long-term tracking process of the algorithm,this paper proposes a model update strategy combined with outlier detection,and analyzes the relationship between the historical change trend of the response peak and the actual tracking effect to determine whether the tracking is abnormal.At the same time,the model update strategy of confidence analysis is adopted to prevent target model drift and filter template pollution;Finally,this paper designs a target loss re-detection mechanism.The frame difference method is used to quickly locate the target position in the whole image for the situation that the target has been lost,which enhances the algorithm’s robustness in the long-term tracking process.