Research On Multi-object Tracking Algorithm Of Multi-egocentric Video

Multi-target tracking aims to obtain continuous trajectory of multiple moving targets in visual inputs,which has great significance for advanced visual tasks such as subsequent behavior analysis,scene understanding and so on.It also is one of the research hotspots in the field of computer vision.In recent years,with the extensive application of wearable camera devices,the visual research based on egocentric video has attracted the attention of researchers.Due to the uncertainty and time-varying character of camera motion,the motion mode of egocentric video is complex and blocked frequently,which causes great difficulties for accurate and stable target tracking tasks.Multi-egocentric video is videos of a number of different visual angles and different trajectories taken by a number of wearable or handheld cameras in the same scene in synchronization time.Multi-target tracking based on multi-egocentric video can solve the problems in single egocentric video target tracking by using abundant information from multiple-view.In the field of visual navigation and other applications,the research of multi target tracking based on Multi-egocentric video is of great significance.In this paper,we study intensively the key problems of multi-target tracking in multi-egocentric video.In order to solve the questions of uncontinuity on moving occurred in egocentric videos,we carry out research into multi-target tracking based on multi-view geometry constraint.And in order to realize robust tracking in three-dimensional space,we carry out research into the camera tracking algorithm of egocentric motion based on multi-model fusion.And we study multi-egocentric video multi-target tracking method in three-dimensional space based on trajectory reconstruction,and carry out the experiment on multi-egocentric video dataset BJMOT,and use EPLF-campus4 and TUM-Freiburg public datasets to prove that our algorithm solves the problem.The main research work of this paper is as follows:Firstly,we study the multi-target tracking method of multi-egocentric video based on the multi-view geometric constraint.In order to overcome the problem of discontinuity of target motion in multi-egocentric video,this paper firstly solves the object occlusion and loss problem based on the synchronous frame homography constraint between multi-view video frames,then further reconstruct location of the target according to the multi-view object spatial location constraint relations.Finally,we use the kalman filter to construct target motion model to refine the target trajectory.The results of experiment on BJMOT and EPLF-campus4 datasets verify that our algorithm solves the problem of targets tracking and prove the effectiveness of our algorithm in multi-egocentric video target tracking,and is more robust than MDP and CMOT algorithm.Secondly,in order to improve the accuracy and robustness of multi-egocentric video target's three-dimensional space tracking,we study the camera pose tracking algorithm of egocentric video based on multiple model fusion.The algorithm improves the accuracy of 3D pose estimation by multiple models fusion.The algorithm firstly initializes and estimates the pose of egocentric camera based on the method of feature point based on moving target detection.Then it combines the direct method and the model of camera motion in space to improve the accuracy and robustness of egocentric video camera pose estimation.The algorithm is verified on the TUM-Freiburg dataset,and the expected experimental results are obtained,and the accuracy is better than the SVO algorithm.Finally,we study the method of multi-egocentric video targets three-dimensional space tracking based on the reconstruction of moving trajectory.We firstly propose a framework of multi-egocentric video targets tracking in three-dimensional space.Based on the above algorithms,we use the method of three-dimensional reconstruction to obtain the target space relative to 3D coordinates,and then use the position information of multiple cameras,the space coordinates of the target relative to the three-dimensional coordinate transformation in world coordinate system,and through the construction of 3D space motion model,access target three-dimensional trajectory.The experimental results on the BJMOT dataset demonstrate the effectiveness of the proposed algorithm.