Research On Theory And Application Technology Of Particle Swarm Optimization Target Tracking Based On TLD

Target tracking technology has always been a research hotspot in the field of computer vision and image processing,which has important application value in intelligent monitoring,visual navigation,intelligent transportation,human-computer interaction and national defense reconnaissance.Over the past thirty years,a large number of domestic and foreign researchers have been improving target tracking algorithms.However,target tracking technology is still a challenging problem due to the complexity of target information,the randomness of target,background interference and target occlusion.At present,although there are many mature target tracking algorithms,most of them still require specific environment and application scope.In particular,a robust tracking algorithm has not been proposed for a wide range of scenarios.Most of the solutions are confined to specific environments,which still need a lot of research.Color feature and HOG feature are two widely used target features in the target tracking field.This paper combines weighted color histograms and HOG features,and reduces the dimension by PCA,which could improve the algorithm tracking robustness in dealing with the situation of illumination variation and target deformation.Aiming at the situation of target occlusion,the particle swarm optimization(PSO)algorithm is applied to target tracking system to optimize the target similarity function,and the particle swarm optimization target tracking algorithm based on feature fusion is proposed.When particle swarm optimization is applied to target tracking system,the inertia weight adjustment mechanism has some limitations.This paper proposes an improved method,which can significantly improve the algorithm efficiency on the premise of ensuring the tracking accuracy.This paper systematically studies each part of TLD model and designs the particle swarm optimization target tracking algorithm based on TLD algorithm.The tracking module,detection module and learning module of TLD algorithm are improved respectively,which improves the robustness of tracking algorithm in complex situations.The main research contents and contributions of this paper are as follows.1.The basic principles of target tracking system is deeply studied.The structured composition of target matched tracking algorithm is analyzed,which is decomposed into two parts: the target feature extraction and similarity function optimization.2.This paper proposes the target tracking algorithm based on color feature and HOG feature fusion.The algorithm combines weighted color histograms and HOG features,and reduces the dimension by PCA.The target is normalized to a unit circle,and the Epanechnikov kernel function with monotonously decreasing central bulge is selected as the histogram weight,which reduces the contribution of edge pixels to the target feature and the influence of partial occlusion on target tracking.At the same time,the color feature has less dependence on the target direction and shape,which has obvious advantages in the case of target rotation and non-rigid deformation in the tracking process.HOG feature has good stability in the case of illumination variation.The two algorithms are fused to give full play to their respective advantages and improve the algorithm tracking robustness in the case of illumination variation and target deformation.3.The particle swarm optimization algorithm is applied to target tracking system,and the particle swarm optimization target tracking algorithm based on feature fusion is proposed.A large number of researches have shown that PSO is an effective optimization algorithm and has good "multi-peak" search ability.When the target is occluded,the target similarity function will also appear "multi-peak".Combining the two methods can solve the tracking problem of target occlusion.Experimental results show that particle swarm optimization target tracking algorithm can effectively complete the "multi-peak" search and achieve precise target positioning.4.Aiming at the limitation of inertia weight adjustment mechanism in the particle swarm optimization target tracking algorithm based on feature fusion,an improved particle swarm optimization algorithm for target tracking is proposed.Firstly,the parameters of algorithm are initialized and the target region is determined.Secondly,the particle maturity is introduced to improve the method of adjusting inertia weight,which can be adjusted accurately according to the different states of each particle in the population.Then,the position and velocity of particle are updated,and the individual and global optimal solutions of the population are also updated.Finally,the position with the largest particle fitness value is determined as the target.Compared with the particle swarm optimization target tracking algorithm based on adaptive inertia weight adjustment mechanism,the improved algorithm reduces the iterations of particles and significantly improves the algorithm efficiency.The improved algorithm can well deal with the tracking problem of partially occluded targets,which can ensure the stable tracking when the similarity function appears "multi-peak".5.This paper improves the tracking module,detection module and learning module respectively by systematically analyzing each part of TLD algorithm,and proposes a particle swarm optimization target tracking algorithm based on TLD.In this paper,the particle swarm optimization target tracking algorithm based on feature fusion is used to replace the original tracking module in TLD algorithm to enhance the tracking robustness of TLD algorithm in the case of illumination variation,non-rigid deformation,scale change,rotation,and occlusion.Aiming at the detection module with large computational complexity in the TLD algorithm,a classifier that can adaptively adjust the variance threshold is proposed to improve the algorithm accuracy and efficiency.In order to solve the problem of online learning sample updating in learning module,a sample deletion mechanism is introduced.In the tracking process,the number thresholds are set for both positive samples and negative samples in the sample library.When the number of positive samples and negative samples both reach their respective thresholds,the sample deletion mechanism will be activated.Then the image blocks classified into the sample library are graded,and the image blocks with weaker representation ability for both positive and negative samples are deleted.Finally,the positive samples and negative samples in the sample library are matched with the current target,and the samples with low representation ability of the current target are deleted.The improved particle swarm optimization target tracking algorithm based on TLD improves the tracking accuracy and efficiency comprehensively.The experimental results show that the particle swarm optimization target tracking algorithm based on TLD is ranked first in a wide range of scenarios compared with the 11 advanced tracking algorithms.The tracking accuracy is 0.687 and the AUC of success rate reaches 0.488.Compared with the second-ranked DLSSVM algorithm,the tracking accuracy is improved by 1.2%.Compared with the original TLD algorithm,the improved algorithm improves the operational efficiency by 25%.