Research On Object Tracking Algorithm Under Complex Scenes

Since the 21 st century,target tracking technology has been applied in various aspects and become an important topic in the field of computer vision applications.It is mainly used in many fields such as transportation,military affairs,security monitoring,human-computer interaction and medical detection.In particular,the important role in the field of security monitoring has become the focus of attention of various countries.Although many achievements have been achieved in the research and development of target tracking technology.When faced with complicated scenes in the real world,it is still a very challenging task because it is affected by various factors such as target occlusion,background interference,illumination change,appearance similarity and camera movement.Aiming at the various difficult problems of target tracking algorithm in complicated scenes,this paper conducts in-depth research on the existing algorithms,the main tasks are as follows :(1)Aiming at the problems of high efficient feature extracting and dealing with the model drift,a kernel correlation filter tracking algorithm based on saliency detection is proposed.The color feature and the histogram of oriented gradient are weightedly merged,and features weights can be adjusted adaptively.For dealing with the model drift,inspired by biological vision mechanism,target salient region is obtained by sampling in the region through the visual saliency algorithm.In that way,the sampling in the area is carried out to complete the global scope search which avoids local maximum.(2)In order to solve the problem of scale fixed in the kernel correlation filter tracking algorithm,this paper proposes a model based on the key points to update the target scale online,inspired by the TLD tracking algorithm.First,we replace the original grid points withthe key points of the target area,and assign different weights to different key points through the response values.Then,the scale of the target is estimated using the discrete degree and weight of the point set.(3)Aiming at the various challenges often encountered in multi-target tracking algorithms,such as camera's sudden motion,occlusion,false detection and appearance similarity,this chapter propose a step-by-step association framework based on kernel correlation filter(KCF).Firstly,accurate detection results are obtained by utilizing a target detector based on the convolutional neural network.Then,a fast tracker based on KCF algorithm is established for each target by using multiple features of weighted fusion.Next,step-by-step association trajectories through the local-global correlation algorithm,and the online random fern is used to re-detect the target in the case of occlusion.Finally,the scale of the KCF algorithm is adaptively updated by using the associated successful detection information.