| With the development of intelligent city construction,a variety of CCTV(Closed Circuit TV)systems has been widely utilized,and a large number of non-structured video data has been created by those systems.This thesis focuses on the implementation of the intelligent retrieval system of panoramic videos,which includes the research on multi-lens panoramic video-stitching technique,object detection and object re-identification based on cloud computing for performing a fast retrieval of objects in large number of panoramic video data.To address the above key technologies,this thesis has studied the multi-lens panorama video-stitching technology,which proposed the application of hierarchical matching to image registration and refreshes frames based on change detection of the video stitching method;the object of research is collected by detecting moving objects,and the feature extraction and description are executed on detected objects for the classification of moving objects when the video-based object detecting technique is used.In the detection of moving objects,a lot of research work has been taken on the classical ViBe algorithm problems,this thesis has proposed the background smoothness of the ViBe background modeling method;In the process of object retrieval,the research on the re-identification of pedestrians is focused by this thesis,for the purpose of solving the problem of overfitting in the metric learning process,an algorithm based on the ring-push restriction by negative samples is raised by this thesis.The main works and research achievements of this dissertation are summarized as follows:Firstly,in the aspect of multi-lens panoramic video-stitching,to address the hot issue that it is not only easy to generate ghost shadow and stitching line distortion because of video parallax for the stable camera views that have overlapped region,but also can’t meet the real-time processing requirements,a video-stitching method has been raised,which uses hierarchical matching for image registration and refreshes frames based on change detection,the hierarchical matching method has better performance of image registration of complicated scenarios with abundant hierarchies,and it is convenient for using the method of refreshing frames based on the result of change detection of overlapped regions to update the stitching lines based on demands.Therefore,both of the accuracy and efficiency of video-stitching can be assured.The experiment’s results show that the method of image registration based on hierarchical matching is a good way to solve the deep distortion problem caused by the complex depth distribution in the panoramic scene.The refreshing of the stitching lines based on the change detection can effectively eliminate the ghost shadow phenomenon and mismatching phenomenon,the result published in the "Botao He,Shaohua Yu.Parallax-Robust Surveillance Video Stitching,Sensors,2016,16(1): 7~19".Secondly,to address the hot issue on the object detection algorithm,it is easy to generate ghost shadow when target is moving in the initialization of ViBe and the status of target changes from stillness to motion in the ViBe processing,the ghost shadow’s ablation rate is slowly and the target splitting problem occurs easily,an improved background sub-traction method based on ViBe has been raised in this thesis,which is based on the research of the appearance of ghost shadow at the single-image initialization of ViBe algorithm and the assumption that background pixels changes over time is a stable process in a short period of time.The initialization of background model is completed by continuous frames,and both the random refreshing method of ViBe and the historic information of pixels are considered in the updating processes of model when this algorithm is utilized.The experimental results show that at the time that the timeliness is satisfied by the algorithm,the appearance of ghost shadow is eliminated,excellent robustness is performed at stable scenarios and the accuracy of object detection is enhanced,the result published in the " Botao He,Shaohua Yu.An Improved Background Subtraction Method Based on ViBe.7th Chinese Conference on Pattern Recognition(CCPR 2016).2016.Part I".Thirdly,in the process of object retrieval,to address the issue that make the positive sample distance less than the negative sample distance and over constrained the negative sample distance as large as possible in the metric learning methods,the problem of over-fitting occurs.An algorithm based on the ring-push restriction by negative samples is raised by this thesis.This algorithm uses the technique of ring-push restriction,which extracts more essential differences among pedestrians,significantly reduces overfitting derived by the limitation of training samples,the influence of noise and background.Thus the final learned metric has better generalization ability.The experimental results show that the accuracy of the VIPeR,PRID450 S,CUHK01 and Market1501 datasets is improved by 0.82%,3.60%,6.21% and 13.03% respectively for the technique of ring-push restriction to the too large negative sample distance.The effectiveness of the method constraint for the optimization of the metric learning model is proved,the result published in the " Botao He,Shaohua Yu.Ring-push metric learning for person reidentification,Journal of Electronic Imaging,2017,26(3): 033005-1~033005-10". |
