Research And Implementation Of Target Tracking Algorithm Based On Kernel Correlation Filtering

Target tracking technology is an important research topic in the field of computer vision.Its main task is to obtain the location of the target of interest in the video sequence,which has been widely used in industries such as human-computer interaction,video surveillance,behavior analysis,security and robotics.At present,although tracking technology has made considerable progress,there are still many challenges to meet in the actual scenario.The scale changes of the target itself and the interceptionbetween the targets,the change of illumination in the environment and the background interference may lead to tracking drift or even failure.In order to solve the above problems,researchers have conducted deep research.Generally speaking,target tracking can be divided into generated model and discriminant model.For the generated mode,when the target is in a complex scene,the tracking result is not ideal because it only considers how to accurately construct the target model and ignores the surrounding background information.Discriminant model obtain the target and the background of the effective decision boundary by training and updating the classifier.Due to it considers background information,This kinds of algorithms show good performance.As a kind of discriminant model,the tracking method based on correlation filtering has achieved good development in recent years.In this method,sufficient training samples are obtained through cyclic shift and the complexity of solving the problem red uced by the characteristic of the circulant matrix,which obtains a faster tracking speed.In this paper,based on the kernel-basedfiltering algorithm,the main work has the following aspects.(1)Aiming at the low tracking performance of the kernel-based filter tracking algorithm in the case of rapid target movement,scale changes and occlusion,a kernel-based adaptive target tracking algorithm based on convolution feature is proposed.Convolution neural network was used to extract high and low convolution features of two layers.The position-kernel correlation filter was used to calculate the response graph of high and low layers.The Coarse-to-Fine fusion two-layer response graph was used to estimate the target position and learnd the estimator scales of one-dimensional filter scale and updates the filter in real time to achieve adaptive target tracking.(2)In order to solve the problem of model error updating when the target is occluded,a new model updating strategy is proposed.According to the maximum va lue of the response graph and the degree of oscillation,we can judge whether the target is obscured or not.Under the condition that the target is not obstructed,in order to improve the accuracy and real-time performance of the tracking result,we will update the appearance model and scale model of the target online.(3)Based on the improved algorithm using MFC,OpenC v third library and VS development environment to achieve a fast target tracking system.The system can select different data sources and calibrate the initial frame target,and it shows the trace speed in the end of the trace.