Research On Key Technology Of Object Tracking In Intelligent Video Surveillance

Visual tracking is a,hot topic in computer vision community as its wide range of practical applications,such as video surveillance,intelligent traffic monitoring and so on.Although a.plethora of tra.ckers have been proposed in recent years,tracking problem is still very diffieult,due to complex real-world environment such as illumina-tion changes,occlusion,background clutter.For nonrigid object tracking,additional challenge factors,such pose variation,severe deformation etc.must be considered.Therefore,it is challenging and urgent to design an robust and generice visual tracking algorithm.This disserttation carries out in-depth research on monocular camera-based object tracking in complex visual surveillance scenes.For single object tracking,we focus on designing robust observation model in two different,ways,generative model and discriminative model,to overcome the partial occlusion and deformation problems.For multiple object tracking,the focus is mainly oi data association problem and some new algorithms are proposed to resolve the long-term occlusion and similar object disturbing problems.The main contributions of this dissertation are summarized as follows:Firstly,we propose a robust and effieient l1 tracker based on laplacian error dis-tribution and struetured similarity regularization in a partiele filter framework.Tradi-tional sparse representation tracker(i.e.,L1T)is based on Gaussian error distribution assunmption,which leads to the 12-norm reconstruction error fidelity term according to the Maximum Likelihood Estimation(MLE).However,Gaussian error distribution assumption is not appropriate for the sparse noise caused by partially occlusion.To overcome this problelm,we model the error term by laplacian distribution.Meanwhile,in contrast to most existing l1 trackers that handle particles independently,we exploit the dependence relationship between particles and impose the structured similarity regularization on the sparse coefficient set.The customized Inexact Augmented La-grange Method(IALM)is derived to efficiently solve the optimization prolem in batch mode.In addition,we also reveal that the proposed method is related to the robust regression with self-adaptive Huber loss function.Experiments on the largest open benchmark video sequences demonstrate that the proposed tracking method presents better performance than the traditional LIT tracker.Secondly,in order to handle the partial occlusion and deformation problems,a novel regional deep learning tracker is proposed.Deep learning has been successfully applied to visual tracking due to its powerful feature learning characteristic.However,existing deep learning trackers rely on single observation model and focus on the holis-tic representation of the tracking object.When occlusion occurs,the trackers suffer from the contaminated features obtained in occluded areas.In this paper,we propose a regional deep learning tracker that observes the target by multiple sub-regions and each region is observed by a deep learning model.In particular,we devise a stable factor,modeled as a hidden variable of the Factorial Hidden Markov Model,to char-acterize the stability of these sub-models.The stability indicator not only provides a confidence degree for the response score of each model during inference stage,but also determines the online training criteria for each deep learning model.This online training strategy enables the tracker to achieve more accurate local features compared with those fixed training trackers.In addition,to improve the computational efficiency,we exploit the structurized response property of the customized deep learning model to approximate the final tracking results by the weighted Gaussian Mixture Model under the particle filter framework.Qualitative and quantitative evaluations on the recent public benchmark dataset show that our approach can keep tracking steadily in complex environments.Thirdly,a new multiple object tracking algorithm based on kernelized correlation filter is proposed to overcome the similar object disturbing problem.Under tracking-by-detection paradigm,data association is the core problem for multiple object tracking and it relays on the object similar metric,which usually affected by the cluttered background.To overcome this problem,we use kernelized ridge regression model to discriminate between the target and its surrounding object with similar appearance.The template filter of the model is trained with all sub-windows sampled around the detection regions.Furthermore,under the well-established Circulant Matrices Theory,regression problem is solved in frequency domain by Fast Fourier Transform.The similar measurement between every two objects is determined by the correlation degree of the template filter and the candidate association regions.We adopt the cascade association strategy to reducing the negative factor caused by detection algorithm,such as missing detection and false detection.First,the tracklet is formed by the objects between consecutive frames with strong relationship.Then,the motion information for each tracklet is exploited.The similar measurement between the tracklet is determined by both the motion and the appearance information.Finally,the trajectory of the object is obtained by the min-cost flow algorithm.Experiments on standard dataset show that the proposed algorithm is robustness to the long-term occlusion and similar object disturbing situations.Finally,we approach the multiple object tracking problem from the subspace clustering perspective and propose a novel tracking algorithm with spatial-temporal constraint.Under tracking-by-detection pa.radigm,the detection set,with outliers can be regarded as the union of samples collected from multiple subspace and each track-ing target corresponding to an independent subspace.Therefore,traditional multiple object tracking problem can be converted into the self-reprcsentative based subspace clustering problem.Specif-ically,we exploiting the spatial-temporal feature embeded in video sequences and construct the local similar matric and the spatial-temporal ex-clusion matric,which is used as the regular expression to constraint the representation coefficient.Then,the object similar measurement is determined by the representation coefficient.The final data association problem is solved by the agglomerate hierar-chical clustering with spatial-temporal constraint.The model is based on the global data representation and it gives our framework the natural ability to handle long-term occlusion and missing detection.The overall evaluation on numerous test video se-quences demonstrate that the proposed tracker perfrmance well under the complex environments.