Research On Spatial Scale Variation With Improved Algorithm Based On KCF For Object Tracking

Humans can get the corresponding information of the pictures by observing them.However,there are millions of video and pictures in the internet,the information of them is mixed,and manual supervising of dirty information costs a lot of manpower and financial resources.Therefore,object tracking algorithm combined with computer vision came into being.This paper aims to track object from video and achieve uninterrupted tracking.So that we can better solves some problem in real life by tracking the information from video.These problems are tracking criminals in the surveillance video,detecting regional intruders,etc.The occlusion,deformation,and blurring of video object tracking have been around for a long time,and many scholars have proposed many excellent tracking algorithms.In this paper,we briefly summarized and analyzed the development process of several tracking algorithms.And then the KCF(Kernelized Correlation Filters)algorithm,which is fast,is selected to improve.Most previous methods use several fully connected layers as regressions to predict the offset of the object positioned frame.This paper introduces a new target detection framework--Grid RCNN(Region-Convolutional Neural Networks),which replaces the typical regression formula with a full convolutional network to predict the position of predefined grid points.And combines with grid points feature fusion and extended region mapping,we can determine the accurate object bounding box.Therefore the accuracy of the object bounding box positioning can be improved.Thus solving the occlusion problem and scale variation in the traditional KCF tracking algorithm.At the same time,in the object model training process,this paper introduces a strategy of updating the model with high confidence.The parameter used in this strategy is APCH,which is the ratio of APCE(Average peak-to Correlation Energy).APCH makes us set the threshold of updating the object model easier.This strategy makes the object model obtained by training better.Meanwhile,in this paper,we compare the performance of the improved KCF algorithm that improved by neural network Grid R-CNN with 9 tracking algorithms.In the forth part of the paper,we introduce the performance evaluation criteria of OPE(One-pass Evaluation),SRE(Spatial Robustness Evaluation).So we can compare our improved KCF algorithm with the location tracking algorithms.The tracking accuracy and learning rate under the same dataset are compared,and the application value of the improved KCF algorithm in the present network environment is analyzed.For real-time tracking problems and automatic learning problems,the improved KCF algorithm that improved by neural network Grid R-CNN uses the Siamese architecture automatic learning optimization function to achieve better CF(Correlation Filters)-based object tracking.This is an innovation in the research of object locating,which provides a more intelligent monitoring and tracking method for future network security,and provides a new idea for the intelligentialize of cyberspace security.