Research And Implementation Of Video Object Tracking Algorithm In Complex Scenarios

Artificial intelligence has seen rapid development in recent years,especially in the field of computer vision.Video object tracking is one of the important research directions,which can be widely used in many fields such as human-computer interaction,driverless,video surveillance.In the past few decades,great progress has been made in the field of video object tracking,especially in recent years,the single-target tracking method using Correlation Filter technology has achieved good tracking results.This paper describes in detail the details of the Correlation Filter based tracking algorithm,and analyzes the problems that need to be solved.Aiming at the object tracking problem in complex scenarios in practical applications,this paper proposes three algorithms based on Kernelized Correlation Filter trackers.The main work of this paper is as follows:(1)A video object tracking algorithm based on complementary feature adaptive selection is proposed.Object feature extraction is a very important part of the tracking framework.This paper analyzes the shortcomings of using only one-channel grayscale features in CSK(Circulant Structure with Kernels)tracker.FHOG(Fast Histogram of Oriented Gradients)feature is robust to challenges such as illumination variation and color changes,as well as motion blur,but is difficult to cope with object deformation and fast motion.In contrast,the CN(Color Names)feature is robust to object deformation and scale changes.Therefore,the FHOG feature and CN feature are complementary features.Therefore,based on KCF(Kernelized Correlation Filter)and CN tracker,this paper proposes a video object tracker that adaptively selects two features to better handle the tracking problem under illumination variation or color changes,object deformation,etc.(2)A video object tracking algorithm based on fast scale Correlation Filter and a relocation component is proposed.The traditional Correlation Filter based trackers can't deal with multi-scale problems accurately and quickly.This paper combines a scale Correlation Filter on the original tracking frame to estimate the scale change of the object and improve the accuracy of object tracking.In addition,in the feature extraction module,the feature representation is still relatively simple,and it is necessary to study a more effective feature representation.In this paper,FHOG feature and CN feature are merged,and a novel hand-engineered feature is proposed to enhance the feature representation of the image,thus improving the discriminative performance of the classifier.Finally,the Correlation Filter tracking framework has always had model drift problems caused by factors such as object occlusion.To this end,the tracking confidence index and the relocation component are introduced,which alleviate the occlusion and model drift problems to some extent.(3)A video object tracking algorithm based on background modeling and adaptive model updating is proposed.The background information that can be utilized during the training of the Correlation Filter is very limited.Moreover,the cosine window is introduced to solve the boundary effect problem,which further reduces the background information.This paper introduces global background information based on the DCF(Discriminative Correlation Filter)tracking framework to improve the discriminative ability of the classifier.In addition,the existing Correlation Filter frameworks adopt the model updating method of linear interpolation,which is difficult to adapt to the change of the object,and is easy to cause model pollution and tracking loss.This paper introduces an adaptive model update strategy to adapt to changes in the object and improve the robustness of the tracker.