Scale Adaptive Visual Object Tracking Based On Bounding Box Regression

Visual object tracking has always been an important research direction in the field of computer vision.With the maturity and popularity of video technology,object tracking has attracted more and more attention,and it has a wide range of applications in the fields of automatic driving,intelligent video surveillance,medical diagnosis and military investigation.The object tracking task is very complex,subject to changes in the object form,variable environment,and occlusion.Although a large amount of relevant research work has emerged every year,object tracking is still a very challenging issue.Bounding box regression is one of the basic components in many 2D or 3D computer vision tasks.Object localization,multiobject detection,and instance segmentation tasks all rely on accurate bounding box regression.This paper explores the object tracking technology based on bounding box regression.The main research work and innovative points are as follows: In order to solve the problem of frequent size changes in object tracking,a scale tracking method based on Weighted Multichannel Bounding Box Regression is proposed.Data augmentation is performed through multi-size sampling.We train a bounding box regressor that can adapt to the deformation of the object,combined with the localization function of Correlation Filter to achieve more refined object tracking.To make better use of multi-channel features,the original singlechannel bounding box regression theory is extended to multi-channel situations.A novel weight function of training samples is proposed,and the time prior information is used more reasonably.The Weighted Multi-channel Bounding Box Regression theory is introduced based on the proposed weight function.At the same time,a measure to measure the diversity of multiple sampling strategies,namely Size Difference Degree,is proposed.Based on this measure,the influence of sampling strategy diversity on tracking performance is studied.The experimental results show that the tracking performance of this method exceeds the mainstream scale tracking algorithm for the interference caused by object deformation,and the comprehensive accuracy(Area Under the Curve,AUC)is up to 4.4%.Compared with the original bounding box regression method,the AUC of the multi-channel bounding box regression method proposed in this paper is improved by 65.4%.The Weighted Multichannel Bounding Box Regression method based on the sample weights proposed in this paper increases the AUC of 27.2% and 5.4%,respectively,compared with the weighting method using only the recent information weight and initial information weight.In addition,the experimental data verified that the size difference degree can effectively measure the diversity of multi-size sampling strategies,which indicates that the tracking performance is positively correlated with the diversity of sampling strategies.