The Research Of Object Tracking Algorithm Based On Feature Fusion

With the rapid development of artificial intelligence and computer vision,target tracking technology is widely used in intelligent transportation,video surveillance.The effect of target tracking is influenced by factors such as background modeling methods,feature extraction,target detection,and tracking scenes.The visual features are unique and selecting the appropriate features to track the appearance of the target is the key to achieving a robust tracking algorithm with good real-time performance and strong adaptability.Most of the current target tracking algorithms are based on a single feature.A single feature is easily affected by factors such as illumination changes,background clutter,and target deformation and it is impossible to describe the target's overall robustness.In view of the above problems,this paper proposes to use multi-feature fusion to describe the appearance of the target,and study the target tracking algorithm based on multi-feature fusion.The main research work of the paper includes:Firstly,the target tracking algorithm can use different visual features to describe the appearance of the target.In order to fit the theme of this article,the classical target tracking algorithms based on a single representative feature are taken as an example.The mean-shift and CN algorithm based on color features,STC algorithm based on spatio-temporal features,KCF algorithm based on gradient features and Struck algorithm based on gray features are introduced.What's more,there use five groups of test video sequences which are Basketball,Biker,David,Jumping and Woman for experiments.The experimental results show that the tracking effect of mean-shift and CN is good when the attitude of the target is changing and scaling,and the tracking effect is poor when the light changing.STC tracks well when the light changing and the target is blocked,and the tracking effect is poor when the target moves fast.KCF tracks well when light changing,the tracking effect is poor when scaling and pose changing;Struck tracks well when scaling and posture changing,and the tracking is poor when the target is moving rapidly.Secondly,in the research of target tracking algorithm based on multi-feature fusion,this paper summarizes the basic knowledge of the traditional particle filter algorithm and block tracking algorithm based on color features.For color features,it is robust to the target deformation,rotation,etc.However,when the target is blocked or illumination changes,the color histogram distribution changing which causes the target appearance description to be insufficient;the blocking strategy of the block tracking algorithm is suitable for target Faceoccl and illumination changes,but it only extracts features from a single block and ignores the spatio-temporal relationship between adjacent blocks.This paper introduces the traditional tracking algorithm based on color and space-time feature fusion,and uses four groups of test video sequences which are Walking2,Faceoccl,Woman and Girl for experiments.The experimental results show that the traditional tracking algorithm based on color and space-time features fusion has high tracking accuracy and robustness in the complex tracking scenes of Faceoccl and pose change.However,this algorithm has no learning detection mechanism.When the target disappears and shows again,the algorithm can't track the target.And the size of the tracking box is fixed and can't zoom with the target scaling.Thirdly,in the research of multi-scale learning tracking algorithm based on feature fusion,this paper uses the idea of TLD algorithm to combine detection with tracking learning.The proposed algorithm fully utilizes the HOG features to be insensitive to occlusion,illumination change and geometric change,as well as LBP features whose advantages is simple,easy and insensitive,and puts the KCF tracking algorithm which combines the HOG gradient feature with the LBP texture features as the tracking module.Considering that the nearest neighbor classifier used in the TLD algorithm is seriously polluted by noise and low in real time,a classifier based on CNN is proposed as the detection module,which makes full use of the advantages of the fast speed of the P-Net network and improves the real-time performance of the algorithm.Based on six video sequences which are Girl,David,Jumping,BlurFace,Faceoccl and Dudek,the experimental results show that the proposed algorithm has high robustness,stability and real-time performance in light change,occlusion,scaling and so on.When the target is disappeared and shows again,the algorithm can redetect the tracking target;When target scaling,the algorithm tracking box can zoom in accordingly.