Reserch On Object Tracking Method Based On Sparse Representation And Multiple Instance Learning And Its Applications In Video Surveillance

With the rapid development of the transportation system in China,there is a fast expansion of the vision-based transportation monitoring and control system in recent years.The object tracking is a key problem in computer vision field,which plays a bridge role in computer vision processing system.Although the domestic and foreign researchers have proposed many effective tracking algorithms in recent years,the object tracking stills a challenging problem due to difficulties,such as illuminate variations,occlusion,scale and rotation changes of the target.Hence,how to robust tracking object in complex scenes with dramatic appearance changes is stills a difficult problem worth studying.Aiming to improve the tracking performance,this dissertation does a deep research for object tracking based on the wide references among this community.This thesis pays its attentions on pedestrian and vehicle tracking in the intelligent monitoring system,arming to improve the accuracy and robustness of the object tracking algorithm and let the object tracking algorithm meet the demand of the intelligent traffic monitoring system.The mainly study of this thesis is shown as follows:1.Analyzing the multiple instance learning algorithm and sparse representation algorithm with the image classification application.An improved multiple instance classification algorithm based on the framework of multiple instance learning and the sparse coding is proposed.Due to the bag-level space based multiple instance learning classification algorithm always ignore the small target region and contains a large amount of redundant information during feature selection,which may cause the information loss for partial bags and affect the performance of classification,we have proposed an improved multiple instance classification algorithm that fuse the multiple instance learning and the sparse coding.Firstly,according to the characteristics of similar samples can cluster into one class,k-means algorithm is used to construct the visual vocabulary for each class of image.To eliminate redundant information,the negative characteristic of negative samples in negative bags is used to constrain the visual vocabulary.The bag feature vectors for each class of training samples are achieved by computing the similarity between the training sample and the visual vocabulary.Then,sparse coding is used to achieve the dictionary matrix for each type of training samples.Finally,the labels for the test images are predicted by linear combination of the dictionary and coefficients to represent the bag-level features for test images.Experimental results show that the proposed algorithm can better solve the problems in multiple instance image classification and achieve higher classification accuracy compared with the other multiple instance image classification algorithms.2.Studying the object tracking algorithm based on sparse representation,which is a typical generative model based tracking algorithm.An improved tracking method is proposed.Firstly,multiple complementary features are used to describe the instantaneous and stable appearance models of the object.Then,a two-stage sparse-coded method,which takes the spatial neighborhood information of the image patch and the computation burden into consideration,is used to compute the reconstructed object appearance.The reliability of each tracker is measured by the tracking likelihood function of transient and reconstructed appearance models.Finally,the most reliable tracker is obtained by a well established particle filter method.Experimental results on different challenging video sequences,including illumination change,occlusion,scale change and rotation,show that the proposed algorithm performs well with superior tracking accuracy and robustness.3.Studying the object tracking algorithm based on multiple instance learning,which is a typical discriminative model based tracking algorithm.Due to the tracking algorithms based on multiple instance learning framework exist some problems,such as not reasonably distribute the instances weight during training classifiers,huge computation and lack the judgment for object drift or occlusion,a novel improved online weighted multiple instance learning(IWMIL)for visual tracking is proposed by introducing the objectness measurement into multiple instance learning.In the IWMIL algorithm,the importance of each instance contributing to bag probability is evaluated based on the objectness measurement with object properties(superpixel straddling),which can discriminatively treat each of instance and beneficial to improve the tracking accuracy.Meanwhile,a coarse-to-fine sample detection method is employed to instead of the exhaustive search method in multiple instance learning tracking,which is benifical to reduce the computation load.Then,to effectively cope with object appearance changes,an adaptive learning rate,which exploits the maximum classifier score to assign different weights to tracking result and template,is presented to update the classifiers.Furthermore,an object similarity constraint strategy is used to estimate tracking drift or occlusion in order to reduce error accumulation and achieve the goal of drift suppression.Experimental results on challenging sequences,including illumination changes,clutter background,occlusion,scale,posed changes and rotation of the object,show that the proposed method can track the object efficiently with superior tracking accuracy and robustness.4.Due to the tracking algorithm based on multiple instance learning framework with the problems that it is without scale adaptively,lack of target discrimination mechanism and time-consuming in feature description,an improved multiple instance learning tracking algorithm based on compressive sensing and superpixel objectness measure is proposed.The proposed tracking algorithm improves the traditional multiple instance learning tracking algorithm from three aspects: target feature extraction,scale based adaptive tracking adjustment and classifiers parameters updating based on target discrimination mechanism.Firstly,compressive sensing theory is used to reduce the feature dimension in multiple instance learning,thus reducing the computational complexity of the algorithm.Secondly,to solve the problem of scale adaptive in MIL,superpixel objectness measure is utilized to carry out the local scale adaptive adjustment.In addition,to reduce the influence of target appearance changes caused by pose change and occlusion for tracking results,target identification mechanism with variable learning rate is introduced to update the classifiers parameters,which is useful to correctly update the classifiers.The similarity between successive frames is used to determine whether the occlusion or drift is exist during tracking,then the variable learning rate is used to update the classifier parameters.Experimental results show a good tracking performance.5.Aiming to efficiently collect the real time traffic information,we study the background modeling and foreground detection method based on ‘low rand +sparse' decomposition,and an effective vehicle counting system for detecting and tracking vehicles in complex traffic scenes is proposed based on low-rank decomposition and sparse approximation.The proposed algorithm first construct the background model for traffic scenes based on background subtraction method with low-rank decomposition,and detects the moving vehicles with sparse approximation.To improve the accuracy of foreground detection,the group sparsity constraint is introduced into the sparse approximation,which is based on the motion characteristics of the target objects.Meanwhile,it can help for improving the accuracy of foreground detections in complex dynamic scenes.For accurately counting vehicles,an online Kalman filter algorithm is used to track the multiple moving objects based on the results of the foreground detections.Then,a reliable tracklet for each vehicle is built by online multi-object tracking.This is beneficial to accurately count the number of vehicles and avoid double counting.In conclusion,this dissertation does a deep research for object tracking based on sparse representation and multiple instance learning.In this thesis,a series of tracking algorithms based on sparse representation and multiple instance learning are proposed to solve the issues during tracking,such as object deformation,similar appearance interference,occlusion,illumination changes and complex background.The proposed tracking algorithms in this paper show good tracking performance to improve the tracking accuracy and robustness in pedestrian tracking and vehicle tracking under complex traffic monitoring scenes.