Particle Filter Tracking Algorithm Based On Convolutional Neural Network

Target tracking is one of the important research directions of computer vision.It is widely used in aerospace,surveillance,biomedical and other fields.The target tracking task is designed to continuously locate and adaptively annotate a specified target in a video containing illumination variation,occlusion.This thesis analyzes the research status of target tracking algorithm,and studies how to improve the accuracy of particle filter tracking algorithm and how to accelerate the algorithm.The essence of the particle filter tracking algorithm is to use the tracker to sequentially determine the foreground or background of each sampled particle.Therefore,the decision function of the tracker is an important factor determining the tracking performance of the algorithm.In this thesis,a tracking algorithm for deep mutual learning is proposed,which can improve the ability of particle filter tracker to extract target features and discriminate.Firstly,the attention mechanism is used to constrain the sampling particles.By adding the attention mechanism layer after the feature map,the interference features are suppressed,thereby improving the ability of the tracker to extract the target features,and solving the problem that the tracker extracts too much in the feature extraction.Focus on the interference characteristics of the background and similar objects while ignoring the characteristics of the target itself.Secondly,the deep mutual learning method is adopted to improve the generalization ability of the tracker.By introducing Kullback-Leibler(KL)loss in two networks with different initialization parameters to change the probability output distribution of the network,the two networks are jointly trained to improve the network.The generalization ability is complemented by the different discriminative information learned by the two networks,and the ability of the network to discriminate the target features is improved.The above two methods together make the tracking accuracy of the particle filter tracker significantly improved.In order to solve the problem of high cost of particle filter tracking algorithm,this thesis proposes a target tracking algorithm based on region of interest(ROI)alignment sampling,which improves the speed and accuracy of particle filter tracker.By introducing a region of interest(ROI)alignment sampling strategy,the picture layer sampling is changed into feature layer sampling,and the original iterative determination of multiple sampling particle iterations is converted into a one-time method for judging all particles,saving a large amount of calculate resources and time to bring the tracker to 21 frames per second.At the same time,the layer-bylayer knowledge distillation training method is adopted to ensure the accuracy of the tracker after acceleration.The tracker with excellent performance is used as the teacher network,and the tracker is guided by the mean square error layer by layer to guide the accelerated tracker to the target feature after acceleration.The discriminative ability is enhanced to approach the teacher network,optimizing the overall performance of the accelerated particle filter tracker.Finally,the comparison experiments on the OTB2015 and VOT2018 benchmark databases show that the parts of the proposed method contribute to the final tracking effect,which improves the accuracy and speed of the particle filter tracking algorithm.In addition,the algorithm of this paper is compared with other algorithms with an accuracy of 0.913,which is at an advanced level.