Research On Object Tracking In The Non-overlapping Multi-camera Views

With the rapid development of security technology,the number of cameras in the living places is increasing,and the monitoring area is gradually expanding.So,the increasing scholars are more concerned about the problem of tracking object in the multi-camera field of view.As we all known,tracking object in the multiple cameras is different from the single camera.Because the factors such as light,object's posture and camera's attributes in multi-camera views are more complex than in the single camera.Especially,in the non-overlapping field of view,the same moving objects may be caught in different cameras,and the characteristics of moving object are discrete both in space and time,which brings a great challenge for object tracking in multi-camera.Thus,in this thesis,we focus on the problem of object tracking in the non-overlapping multi-camera field of view.1)To track objects in the multi-camera field of view,we need to research the method of object detection and tracking in the single camera.To improve the accuracy of detection,a Mixed Gaussian Model method based on wavelet transform is proposed.In this method,the wavelet transform is used to remove the noise and make the useful information clearer.Then,the Mixed Gaussian Model is established to extract object.The experiments show that our method is more accurate than other method.In the stage of tracking,because the TLD is poor in real-time and sensitive to illumination,we propose a TLD method based on Kalman filter.Firstly,the Kalman filter is used as the tracker for the TLD instead of optical flow,and the result of the detector in TLD is applied to update the tracker.Finally,the results of both tracker and detector are integrated to locate the moving object.Experiments show that our method is more real-time and accurate than TLD algorithm.2)To reduce the time and space differences caused by blind spots between cameras,a topological estimation method based on Gaussian and cross-correlation functions is designed.In this method,the results of the object detection and tracking in the single camera are used to acquire the topological nodes.And the connections between the nodes are estimated according to the average cross-correlation functions in more than one time periods.Finally,we adopt the Gaussian model to describe the probability model of transfer time on each connection,making the result more stable.This method can accurately estimate the topology of the camera network without obtaining the connection relationship between the cameras in advance.3)Firstly,to reduce the influence of illuminate between cameras,an object matching method based on cumulative color characteristic transfer(CCCT)models is proposed.In this method,many images is applied to learn the CCCT models.With the CCCT models,the object image is transferred in the field of view from a camera to anther camera,which makes the object matching between cameras more easily and more accurate than other methods.Secondly,in order to measure the object's similarity between the cameras,an object matching method based on SIFT is put forward.At first,the SIFT feature of object is extracted to building the apparent model.And an object matching strategy based on sequence is proposed to match objects between cameras.The experiments show that our method can accurately identify the object between cameras.4)To track objects in the non-overlapping multi-camera views,a new object association method that combining the topology and apparent model is proposed to track objects in the non-overlapping multi-camera field of view.In this method,we integrate the topology,apparent model and matching strategy into the framework of object association method.Firstly,the candidate objects are decided according to the topology of the camera network,and the CCCT models is used to reduce the influence of illuminate between cameras.Then,the apparent model of object is extracted to matching the objects between cameras.With this method,we can reduce the computational complexity and improve the accuracy in the process of object association.