The Research Of Multi-camera Pedestrian Tracking Algorithm Across Non-overlapping Views

As the development of society,persons' demand for security is growing,which makes the Video Surveillance System widely distribute in all kinds of places.In the widely distributed Video Surveillance System,it is unpractical to recognize the suspicious targets and realize continuous tracking.Under the promotion of social demands,object tracking and handoff technologies achieve extensive research.To overcome conditions that include illumination variance,object deformation,occlusion and so on,a host of researchers propose many methods for object tracking and handoff.In object tracking,online-boosting based object tracking method treats the tracking problem as a binary classification issue.And the correlation filter based object tracking methods utilize the correlation to realize tracking.The Minimum Output Sum of Squared Error Filter(MOSSE)and the Kernelized Correlation Filter(KCF)are two representative filters.In this paper,online-boosting based object tracking method and MOSSE filter based object tracking method are firstly researched.Then an effective strategy is proposed to improve the performance of kernelized correlation filter based tracking method.In MOSSE filter aspect,the original filter which is based on gray template is researched in advance.Afterward HOG feature based MOSSE filter is proposed for tracking which replaces the gray template with HOG feature.In kernelized correlation filter aspect,it's tracking performance is influenced by the following factors: firstly,object deformation,rotation and other conditions would destroy the circulant matrix of samples which is the key of kernelized correlation filter;secondly,when the object moves rapidly or becomes sheltered by something,the actual tracked object would not be contained in the candidate set which is produced by cyclic shift.This paper introduces a histogram redetection strategy into KCF and applies an effective scale estimation method for overcoming the factors mentioned above.The histogram redetection strategy and scale estimation method improve the performance of original KCF.Finally,OTB50 and VOT2014 are utilized to verify the effectiveness of our proposed method.The experiments demonstrate that the modified method outperforms the original KCF on OPE,TRE and SRE evaluations.In object handoff,a host of researchers focus on two directions: appearance model and the mixed model combining appearance and spatiotemporal constraint relationship.The appearance model based method can also be called Object Re-identification.In actual environment,many cameras discretely distribute in vital positions.However,there still exists monitoring blind area,producing non-overlapping views.Because the area out of the camera views is also large and complicated,the spatiotemporal constraint relationship is uncertain.And diverse Video Surveillance Systems own different spatiotemporal constraint relationships.Therefore,the mixed model based method loses the generality.Among the appearance based methods,the adaptive learning methods including the object model learning and the metric learning require training samples which also results in losing generality.Therefore,the feature representation of objects is researched to build object model and distance metric function is used to calculate distances between different objects.Then the object handoff among different cameras with non-overlapping views is accomplished by the model and distance metric.Consequently,the object would be tracked continuously in Video Surveillance System.In this paper,the histogram in log-space,covariance descriptor and Hierarchical Gaussian Descriptor(GOG)are combined to build object model.In the procedure of building object model and recognizing objects,the person object is segmented into three parts including head,torso and leg for extracting features.Then,the distance between pedestrians is calculated by assigning different weights to corresponding parts.Finally,the same person would be re-identified among diverse cameras with non-overlapping views.The experiments in VIPeR dataset demonstrate the effectiveness of the proposed method in this paper.The experiments demonstrate that compared with SDALF and ELF,our proposed method achieves higher recognition rate.