Research And Application Of Visual Tracking Based On Deep Feature And Correlation Filter

UAV visual tracking has an urgent demand for the technology in the field of computer vision,in which target tracking is its basic research problem.Visual tracking mainly tracks the target and records the motion trajectory of the target through the target sample in the first frame.Compared with traditional target tracking,UAV target tracking has more special scenes and higher requirements for algorithms.The accuracy of target tracking can be improved after more than 20 years of research on the accuracy of target tracking and rotation.Through the research and analysis of target tracking problem and aiming at the shortcomings of existing algorithms,this paper puts forward three solutions.The main work of this paper is as follows:(1)A combined feature is proposed to realize the complementarity of multiple features,and improve the expression of target features in the scene of deformation and illumination change.We use the convolution feature extracted from vgg-19 model as the deep feature and the color histogram feature as the shallow feature to learn the correlation filtering response of the two features respectively,and reasonably fuse the two features to form a combined feature.Therefore,we design a dynamic weight strategy to sort and screen the response of the deep and shallow features according to the size,and select the state with the lowest overall loss through calculation The fusion parameters of shallow features are calculated.(2)A reliability judgment mechanism based on the average peak correlation energy response is proposed,which solves the problem that the traditional algorithm is easy to introduce clutter into the tracking model when there are scenes such as occlusion,rotation and deformation,resulting in error accumulation and final tracking failure.We use the average peak correlation energy to calculate the reliability of the tracking results.The average peak correlation energy response can reflect the confidence of the tracking results,so as to dynamically reflect the reliability of the tracking results.We only use the tracking results whose average peak correlation energy response of the tracking results is higher than the threshold to update the model,which effectively avoids the introduction of errors.(3)A hierarchical update strategy is proposed.According to the characteristics of UAV target tracking scene,we introduce the context information around the target to improve the recognition ability of the model to the background.In this regard,we design a new tracking model updating strategy,extract the three-layer convolution features and color histogram features of the target,as well as the depth features of the target context and background information,add the depth features of the target and the context and background information to train the target model of the target’s deep features,and train the target’s color histogram features to obtain the target model of the target’s shallow features,When the average peak correlation energy response is greater than the threshold,the deep and shallow models of the target are updated respectively.