Research On Correlation Filtering Tracking Algorithm

Target tracking is one of the important research directions in the field of computer vision.It has received extensive attention in both academic and application fields.The main applications are intelligent video surveillance,human-computer interaction,and automatic driving.The main purpose of the target tracking task is to give the initial position of the target and track the specific location of the target in real time in a continuous video sequence.After years of research and development,the performance of the tracking algorithm is still affected by factors such as rotation,rapid motion,occlusion and scale changes.This paper studies the basis of the kernel-related filtering target tracking algorithm.Firstly,this paper uses the attention mechanism and occlusion detection to improve the kernel correlation filtering target tracking algorithm.The convolutional neural network is used to extract the convolution feature,and the convolution feature is used to train the kernel correlation filter.Since the convolution feature contains rich semantic information,it is more favorable for the filter to distinguish the target from the background.Using the cross-correlation matrix of two samples to calculate the attention weight,combined with the attention weight and the kernel correlation filter,the method introduces the attention mechanism into the kernel correlation filter,and finds the scale by the cross-correlation matrix of two samples.Factors such as changes in variation,noise,and light intensity affect smaller key features and give high-priority values to key features,thereby enhancing the judgment of nuclear correlation filtering.In order to prevent the model from being updated when the target is occluded,the noise is continuously accumulated.Two classifiers are used to detect the target and the background respectively,and the occlusion is determined according to the response values of the two classifiers,and then the model is adaptively updated.Then,this paper proposes a kernel-correlated filtering target tracking algorithm based on Gaussian mixture model.The convolutional neural network is used to extract the convolutional features and establish a Gaussian mixture model of the target appearance,to make up for the deficiency of the kernel correlation filter in the appearance model,and to update the Gaussian mixture model using an efficient online update method.The multi-scale-multi-shape tracking method is used to accurately estimate the target scale and shape after the initial tracking,which makes up for the deficiency of the scale pyramid method and has lower computational complexity than the scale pyramid method.Finally,this paper combined siamese network and deformable convolutional network,and proposes a deformable siamese network.The network uses the target samples and search regions as inputs,and uses a convolutional neural network to predict the offset of the deformable convolution.The convolution feature is then extracted using a deformable convolution.Finally,on the convolution feature,the correlation filter is used to detect the target.The above three algorithms are quantitatively and qualitatively analyzed on the public dataset.The experimental results show that the proposed algorithm effectively improves the robustness and accuracy of the algorithm under occlusion,rotation and deformation.