Research On Target Tracking Method Based On Correlation Filtering

Machine vision is widely used in industry,military and aerospace industries,and target tracking is an important application direction.The target tracking algorithm achieve better tracking effect for both single target and multi-target tracking in complex scenes.Correlation filtering tracking algorithms stand out from many target tracking algorithms due to their high accuracy,fast tracking rate,and strong robustness.However,tracking results in practical applications are easily affected by the performance of the algorithm and the surrounding complex environment,therefore it is of great research significance to improve the performance of the tracker in the complex environment.In order to solve the problems of occlusion,scale change,rotation and background clutter in the process of target tracking,the main work and contents of this thesis are as follows:(1)To solve the problems of bad manual feature tracking accuracy and slow depth feature tracking speed,this thesis proposes a target tracking algorithm based on feature fusion.First of all,this thesis combines the HOG feature,CN feature and depth feature,the purpose is to improve the ability of the tracker to distinguish the target,and the tracker can achieve a balance between speed and accuracy.In order to solve the scale transformation problem,the scale information of the target is extracted by the scale filter,and then the position information of the target is combined with the scale size to improve the tracking accuracy and success rate.(2)In order to solve the commom problems of occlusion,rotation,background clutter in target tracking,a target tracking algorithm with context-aware information and re-detection mechanism is proposed.First of all,this thesis combines the HOG feature and the color feature in order to solve the problems of deformation and motion blur in the tracking process.Then,the context-aware framework is introduced,and four background information of the target is fed into the filter,with the purpose of estimating the motion state of the current target.Finally,the reliability of the tracking results is estimated by using the peak sidelobe ratio and the re-detection mechanism.(3)The two improvement strategies proposed in this thesis,and several algorithms and different video sequences are selected for comparative experiments.The first improvement strategy in this paper: the success rate is increased by 0.116,and the precision is increased by 0.231.And the second improvement strategy: the success rate is increased by 0.212,and the precision is increased by 0.301.It proves the feasibility of the improvement strategy proposed in this paper.