Correlation Filtering Based Object Tracking With Spatial Information

Object tracking refers to the method of learning and training the model from a series of continuous video images to obtain the latest motion information of the target,and locate the selected target accurately in real-time.Object tracking originated in the military field and was used in military equipment such as radar detection and missile guidance.It is a technology,which mainly simulates of the human visual system,using video cameras and computers instead of human eyes and brains to deal with related video information.With the advanced theoretical basis and technology,it has a wide range of practical applications in many fields.This article conducts in-depth research on background changes,target occlusion,deformation and other issues in the image sequence tracking.The main results are as follows:1.A particle filter object tracking method based on spatial confidence is proposed.This method uses the Monte Carlo framework by sampling randomly near the target in the current frame image to obtain individual particles.Each particle is filtered using a basic filter,then the relevant response score is obtained.According to the distribution of the particle,the peak sidelobe ratio is used to calculate the weight of each particle,and the particles are combined to predict the target position and size according to the weights.Through acquired predictively target position and size information,the correlation filter parameters are updated so that the filter adapts to the new target and the next frame can be tracked.2.A multi-scale tracking method based on kernelized correlation filtering is proposed.The method uses the cyclic shift of the circulant matrix to sample the positive and negative samples in the image,which makes full use of the circulant matrix to achieve diagonalization in the frequency domain.It converts matrix operations,which is extremely time-consuming,into point operations between elements in the vector.This method greatly improves the computational efficiency and enables real-time tracking of the target.A scalar filter is introduced to prejudge the size of the target,which can solve the problems of scale change and further improve the accuracy of object tracking.3.A multi-scale tracking method based on convolutional neural network is proposed.This method extracts convolutional neural network in the image.Meanwhile,it tracks each target through a weak tracker to obtain the tracking position of each layer,and combines all the weak trackers according to different weights to a strong tracker.Then we can use the strong tracker to accurately predict the position of the target in current frame and update the weight information for each weak tracker.Next,an independent scale filter is used to scale the target location,and the target size is determined according to the scale information of the target.Finally,the parameters of the weak tracker model and the scale filter model are updated respectively.The method has good robustness on illumination and occlusion.