Research On Relative Motion Network And Bayesian Probabilistic Estimation Based Multi-target Tracking Method

Vision based multi-target tracking is a hot topic in the field of computer vision,which has widely used in the fields of video surveillance,automatic driving assistance,mobile robot navigation and traffic safety.Due to many factors,such as the complexity of the environment,the targets5 appearance variation,target number variation and occlusion et al,visual multi-target tracking becomes very interesting but challenging.Especially in the moving platform,the motion of the camera itself leads to the change of background,which makes the moving targets and the background hard to distinguish,and increases the complexity of the multi-target tracking problem.To address the multi-target tracking issues with a moving camera,a Bayes posterior estimation method is proposed in the framework of tracking-by-detection.The main contents include:(1)The method of camera motion estimation is addressed.This paper estimate the camera motion by estimating the stationary geometric features,the corresponding relationship between the targets in 3D space and in 2D image plane is established at each moment's camera state,which is prepared for computing the target motion model and the observation likelihood.(2)The method of fusing multiple different detection cues to adapt to the change of target appearance and complex environment were explored.The detection cues are provided by deformable part model(DPM)and a color-based MeanShift tracker,which allows the addition or deletion of detection cues as needed.(3)The relative motion network in 3D space is proposed to solve the track fragments and tracking drift due to missed detection and occlusion.Relative motion network describes the 3D relative movement between targets,which is more accurate and intuitive than the 2D motion network in image plane.It also can adapt to complex motion of target and reduce the dependence of the tracking algorithm on the detection results.(4)Reversible Jump Markov Chain Monte Carlo(RJMCMC)particle filtering is exploited to solve the posterior estimation problem of Bayes,it can add or remove target randomly by reversible jump move operation,and it can adapt to random changes in the number of targets.To verify the validity of the proposed method,three benchmark datasets(ETH-Bahnhof,ETH-Linthescher,and ETH-Sunnyday)and some videos collected by ourselves were used.The experiments show that the proposed method is outperformed in handling the track fragments due to occlusion and unreliable detections.