Study On Long-term Target Tracking Algorithm With Correlation Filter

Target tracking refers to positioning of targets in continuous video sequences.It is widely used in many fields,such as security monitoring,localization and mapping,smart city and so on.In actual tracking process,there are many challenging factors,such as background illumination change,target fast movement and nonrigid deformation.On the one hand,when tracking accuracy is low,traditional algorithms often continue to update template,which leads to tracking box drifting and unable to recover positioning.On the other hand,when object is seriously blocked or even left the field of vision temporarily in long-term target tracking,traditional algorithms do not have the ability to retrieve the target again.Therefore,aiming at the above problems,this paper studies target long-term tracking algorithm based on correlation filter.The primary research contents are as follows:(1)To solve the problem of template pollution and inability to evaluate the reliability of tracking results,a target long-term tracking algorithm based on confidence evaluation strategy is proposed.Average peak correlation energy and score of color histogram are taken as the confidence and compared with dynamic threshold.According to evaluation of confidence,tracking template is updated to reduce the interference of complex scenes,such as background clutter,motion blur and occlusion.When reliability of tracking box is low,activate the target re-detection mechanism.Simulation experiments are carried out on OTB-2015 datasets,accuracy and success rate of the algorithm are 85.5% and 79.8%,respectively,showing good robustness.The experimental results show that long-term target tracking algorithm based on confidence evaluation can effectively protect the template and enhance capacity of resisting disturbance.(2)Aiming at the problem that target can not be located once more when it is lost,a long-term tracking algorithm based on re-detection is proposed.When confidence score is lower than global detection threshold,the improved EdgeBoxes algorithm is used to detect candidate boxes quickly.On the contrary,the improved sparse coding method is used to extract candidate boxes.In the OTB-2015 datasets,the accuracy,success rate and average running speed of the algorithm are 86.1%,80.4% and 35.7FPS,respectively.Experimental results show that the long-term target tracking algorithm based on re-detection can track stably in occlusion,out of view and other video sequences,and ensure real-time running of the algorithm.