Research On Target Tracking Algorithm In Complex Scenes

As a popular research direction in the field of computer vision,video target tracking has been widely applied to the land/sea/air transportation,public security,civil security and so on.However,reliable and accurate target tracking is a challenging task due to many factors in practical scene,such as the light variation,the background and clutter,the overlapping and the occlusion between objects,the camera perspective transformation,and so on.The goal of visual target tracking is to locate the target in the collected video sequence,and obtain the important information about the target state such as coordinates of position,size,velocity,contour,as well as the movement path.The result of target tracking is essential for the following advanced video analysis and processing.To address the shortcomings of the traditional particle filter method in video tracking,two improved algorithms are proposed in this thesis under tracking by detection framework.The proposed algorithms combine the Adaboost detector with the particle filter,and are applied to the single target tracking and multi-target tracking respectively.The main contributions of this thesis are as follows.(1)For traditional particle filtering,template updating will generate error,leading to tracking drift or even failure,and particle filter is unable to be reinitialized.To deal with the issue,a single target tracking method is presented by combining online detection with particle filtering.Specifically,online Adaboost is adopted to extract feature and detect the object.On the other hand,the sample set is updated according to the tracking results of particle filter,and the newly updated sample set is utilized to classifier training and detection.By comparing the output of both the detector and tracker,the result with higher confidence is chosen as the final target position.The simulation experiment shows that the proposed method can effectively overcome the drift problem caused by the frequent entries and exits of target and the template ambiguity,thus improve the tracking performance.(2)The common mixture particle filter is capable for multi-object tracking when the number of targets is fixed in the scene.Furthermore,the overlapping among targets will result in the ambiguity of the target,and the mixture particle filter can not be automatically initialized.To resolve these problems,an algorithm is designed that can detect and track simultaneously.More precisely,a cascade Adaboost detector is introduced to detect the entries and exits of targets,and the proposed sampling distribution is constructed by the detector's results.At the same time,the HSV histogram and HOG feature are combined to represent the object appearance more efficiently.Finally,the multi object tracking can be achieved under the framework of mixture particle filter..Experimental results show that the proposed method can not only automatically track targets with variable number,but deal with the ambiguity for targets with similar appearance.A robust multiple target tracking is realized.