Correlation Filter Visual Tracking Based On Multiple Peaks Detection In Reality

Visual object tracking is one of the basic research directions in computer vision.In recent years,with the rapid spread of high-performance computers and video equipment,people are paying more and more attention to visual object tracking.Until now,a variety of visual tracking algorithms have been proposed.Among them,the best performance and the most researched are deep learning based object tracking algorithm and correlation filtering based object tracking algorithm.The object tracking algorithm based on correlation filtering has attracted a large number of researchers at the beginning of its application with its characteristics of tracking performance and speed.However,the current object tracking algorithm based on correlation filtering still cannot track the object well,especially in complex tracking scenarios.At the same time,because the object tracking algorithm based on the correlation filtering needs to cyclically shift the sample,this process will lead to boundary effects,which limits the performance of the tracking algorithm to some extent.The BACF algorithm is a object tracking algorithm based on correlation filtering.The algorithm solves the boundary effect problem better,and has better tracking effect in multiple tracking scenarios,and the tracking speed is faster.However,the tracking result of the algorithm is not ideal for tracking disturbances such as background clutter,similar objects and illumination variation,or motion blur and deformation of the tracking object.At the same time,it is more complicated for the tracking object to be occluded.In this case,the algorithm cannot continuously track the object or even track the object.Therefore,this work will based on the BACF algorithm to conduct a research on these problems,and the main research contents are as follows:(1)Aiming at the problem that the BACF algorithm has poor anti-interference when the background clutter,similar objects and illumination variation occur in the tracking scene,the feature fusion strategy is used to enhance the robustness of the tracking algorithm.Firstly,the model updating method of BACF algorithm is improved by linear interpolation.Then a multicorrelation filtering feature fusion strategy is adopted to feature fusion of HOG features and CN features with complementary characteristics to enhance the anti-interference ability of tracking algorithm;(2)In the tracking scene for one of the target occlusions,the response graph output by the BACF algorithm often contains multiple locations with large response values,and the algorithm cannot accurately locate the tracking target.The strategy of forward multi-peak detection is used to obtain the tracking target confidence at multiple locations to accurately determine the location of the tracking target.Firstly,the peak neighbor ratio method is used to discriminate the ambiguous response graph.Then,based on the target confidence criterion,the tracking target confidence degree at the candidate position of the tracking target is obtained,and finally the tracking target is accurately positioned;(3)In the tracking process,the tracking target size constantly changes,especially when the tracking target aspect ratio changes,the BACF algorithm cannot accurately estimate the tracking target size.Using the strategy of tracking the length and width of the object respectively,firstly,the larger search area is used to estimate the exact position of the tracking target,and then the long and wide Gaussian pyramid is used to estimate the optimal scale of the length and width of the tracking target;(4)In order to verify the effectiveness of the improved BACF algorithm,this thesis selects the OTB benchmark set as the verification platform of the algorithm performance,and verifies the performance of the improved algorithm from qualitative and quantitative aspects.The first improved algorithm improved 3.2%accuracy and 2.3%coverage on the OTB-2013 dataset;the second improved algorithm improved 3.9%accuracy and 3.7%overlap on the OTB-2015 dataset.This verifies the effectiveness of the improved algorithm and enhances the robustness of the tracking algorithm.The improved algorithm runs on the Intel i7-4790K CPU computer with tracking speeds of 26.3FPS and 27.3FPS,respectively,meeting real-time requirements.