Research Of Visual Tracking Based On Correlation Filter

Visual tracking is a research hotspot in the field of computer vision,and has important research value in many fields such as military and security.In recent years,the correlation filter-based tracker attracts great attention due to its robust and fast tracking performance.It would make the tracking task more difficult due to these challenging factors such as deformation,background blur and other factors in complex environment.In addition,the performance of the monitoring equipment is limited,resulting in the obtained apparent information not robust,which seriously restricts the performance of the tracking algorithm.In order to solve these problems,we proposed two algorithms:(1)A correlation filter tracking algorithm based on global context and feature dimensionality reduction is proposed.The image patches uniformly around the object are extracted as negative samples to make full use of the background information,and thus the similar background patches around the object are suppressed.In order to solve the problem that the HOG feature is sensitive to deformation,the color feature is added to represent the moving object,and the adaptive fusion of the two features is computed at the level of the response graph.Finally,the scale pyramid is constructed at the determined object position,and a correlation filter is separately trained based on the principal component analysis to solve the problem of the change of the object scale in the tracking.The algorithm is verified and analyzed on the OTB standard dataset.The results show that the proposed algorithm is more robust and can effectively cope with the deformation,background interference and scale variation.(2)A correlation filter tracking algorithm based on multi-view detection and multi-template learning is proposed.In view of the inability to learn effectively due to excessive gradient changes during network training,the OIM loss function trains the convolutional neural network through a non-parametric optimization to obtain more discriminative intra-class features.And multi-view matching strategy is proposed where the perspective is used to detect the object,and the different views are explored to overcome the interference of similar pedestrians.The multi-template learning is introduced in the tracking module for constructing a multi-pose apparent model to improve the description ability of the object,and is helpful for determining the object location more accurately.In addition,a re-detection mechanism is added to solve the object tracking drift problem.Finally,the videos in the actual monitoring scenario are collected and tested.The results show that the proposed algorithm effectively overcomes the limitation of monitoring equipment in the actual monitoring scenario,and achieves the purpose of long-term effective tracking under the monitoring scenario.At the same time,the proposed algorithm is compared with other algorithms such as ROT and fDSST to verify the effectiveness of the proposed algorithm.