Multi-person Tracking Based On Sparse Representations Of Features

The computer and information technology has developed rapidly in recent years.Accompanied by the growth of image,video and other information data,the development of computer vision and artificial intelligence has been promoted.As a hot research topic in the field of computer vision,video object tracking has a good prospect in the application of intelligent video surveillance,human-computer interaction and intelligent security.The main task of video object tracking is to locate the interested target continuously and accurately in the video.At present,there are some achievements in the field of video object tracking.However,there is still a distance between video object tracking and practical application due to the complex and changeable environments,the inevitable interactions and occlusions between targets,the scale changes of targets and so on.Therefore,it remains a great research value in video object tracking technology.As the main part of society,human is an important research object of video object tracking.The research topic of this paper is multi-person tracking,because there are more than one pedestrian in specific video scenes.Sparse representations used as the target descriptor is robust to partial occlusions.Therefore,sparse representations is utilized to describe the target in this paper,and an online multi-person tracking algorithm based on sparse representations of features is introduced.In order to discriminate the target and background better,a classifier based on sparse representations is constructed for each target.Then,the Bayesian inference is employed for pedestrian tracking by the classifier,and the optimal estimation of the target state is viewed as the output of the tracking result.Finally,an integrated framework is proposed to realize multi-person tracking by combining several single-object trackers.Based on the constructed over-complete dictionary,the sparse coefficients of the combined features(gray feature,HOG feature and LBP feature)are utilized for the description of the object.Tracking is to locate the pedestrian continuously from the pedestrian appearing to disappearing in the scene.The appearance model which is constituted by the over-complete dictionary and the classifier should be created for each target.When the new image coming,the Bayesian inference is employed to estimate the optimal state of the target.For each target,every target is corresponding to an independent tracker.Multi-person tracking is our main work,so an integrated framework is provided to combine multiple single-target trackers.Under the proposed framework,we mainly determine the beginning and end of single-target trackers,and relate the pedestrians in different frames.The further judgment of beginning and end of the independent tracker is based on the detections,and the association is to solve the problem of data association between detections and multiple tracking results.To verify the effectiveness of the proposed algorithm,we validate the three standard datasets on PETS09 S2L1,Town Center and Parking Lot,respectively.The experimental results show the robustness performance of the proposed multi-person tracking algorithm based on sparse representations of features.