Research On Correlation Filter Tracking Algorithm Based On Historical Information

As a hot research in the field of computer vision,object tracking algorithms have been closely followed by researchers in related fields due to their wide application scenarios.With the continuous development of intelligent hardware and Io T(Internet of Things)devices,object tracking technology is widely used in the fields of drone positioning,intelligent security,human-computer interaction,unmanned driving,and intelligent logistics.In recent years,especially after the correlation filter was proposed,a large number of excellent algorithms have been proposed in the field of object tracking field.While the object tracking algorithm gradually ensures real-time online tracking,the tracking accuracy and precision are also increasing.However,objects are often faced with complex realistic scenes such as illumination changes,background changes,size changes,rapid motions,occlusions,and similar interferences during the movement process.This provides some challenges to develop the real-time algorithms with high robustness and better generalization capabilities.In order to solve the issues caused by target occlusion,fast motion,model update in the tracking process,this paper proposes three improved correlation filter tracking algorithms.The comparison of experiments with mainstream algorithms on public data sets shows the effectiveness of the improved scheme proposed in this paper.The main contributions are given as follows:For the situations where the target is occluded or changed abnormally,this paper proposes a method to evaluate the tracking effect based on the response map of each frame of image.And on the basis of this evaluation method,the advantages of different tracking algorithms are combined to achieve a better tracking effect.In the model update stage,the paper realizes the dynamic adjustment of the learning rate of the model update based on the realistic effect of the current frame response map.Experiments on the OTB-50 and OTB-100 data sets show that the above optimization method based on response map has achieved good results.In the process of model updating,mainstream object tracking algorithms usually give the historical frame and the current frame model a fixed learning rate,and superimpose them as the tracking model for the next frame.Through analysis and demonstration,it is found that historical frames will continuously multiply the learning rate,which makes the learning rate show an exponential decay trend and is not good for the memory of historical information.In this paper,the gradient of the adjacent frame model is introduced into the correlation filter tracking algorithm,and the first and second moments of the gradient are used to construct the memory-capable and variable model update coefficients to compensate for the attenuation of the learning rate of historical frames during object tracking.