A Robust Multi-target Visual Tracking Method Using TLD Framework

In the 21 st century,with the rapid development of artificial intelligence and Internet if Things,as the most important way for computers to access the information of nature directly,the research in computer vision area has become a hot spot once again.Visual tracking is one of the main research directions in this field.Today's state-of-the-art methods for visual tracking is to improve tracking by detection.It estimates the position of an object and adjust the parameters of object's model.Zdenek Kalal put forward a new single-target long-term tracking algorithm in 2012.This algorithm combines the traditional detecting method and the traditional tracking method and continuously updates the model of moving target with a semi-supervised learning manner to solve the problem that deformation of the target and partial occlusion may cause tracking failure.It is the main difference between this algorithm and the traditional tracking algorithm and it makes the tracking system more stable,robust and reliable.However,this algorithm also has some limitations.On one hand,since the algorithm needs to select an unknown target from the first frame of the video sequence manually,it cannot automatically identify the target that has been marked in advance or track multiple targets.On the other hand,when the attitude of the target changes a lot,such as pedestrian turning,motor vehicle steering,sitting people standing and walking,it is much likely to cause the tracking failures.This paper proposes a novel multi-target visual tracking method based on the classical TLD framework.The main results of this research are as follows:(1)Proposing an improved cascaded classifier to solve the problem that when the original TLD framework update its online template library,it may cause accumulated error.(2)Adding an auxiliary detector to the original TLD framework to solve the problem that the original framework is sensitive to the deformation and rotation of the tracking target.(3)At last,for the problem that the original TLD framework can only target a single target,adopting the non-category tracking strategy,treating each target as an independent individual and taking parallel processing.Through the redesign of the overall structure and the improvement of each module,a multitarget tracking system based on TLD algorithm is realized.After many experiments,it is proved that the multi-target tracking algorithm based on TLD framework proposed in this paper not only solves the problems caused by temporarily lost and partial occlusion during the movement of moving target without affecting the efficiency of system,but also is robust to the deformation and rotation of moving target.