Research On Multi-object Localization And Correspondence Techniques Based On Homographic Constraint In Overlapping View

In recent years, with the development of intelligent surveillance technology, the traditional single-camera surveillance system has been failed to meet the demand for intelligent monitoring. On the contrary, using multiple collaborative cameras can effectively expand the scope of monitoring, rich moving object information, solve problems such as object occlusion. Therefore, multi-camera surveillance system has been increasingly studied and widely applied. But, some difficulties also need to be solved. How to effectively integrate the rich information coming from multiple cameras is the key to the multi-camera video surveillance system. The focus of this thesis is on the localization and correspondence of multiple objects across overlapping camera views, as well as positioning and matching of multiple objects under moderate severe occlusion in crowded scene. The main contributions and innovation points are as follows:(1)Consider the blindness of the traditional method for homography matrix estimation, a novel method based on planar plane segmentation is designed. It splits the planar plane of interest in which the extraction of correspondence points by SIFT and the estimation of homography matrix with RANSAC are performed then. Experimental results show that this pretreatment greatly reduce the time of SIFT feature point extraction and RANSAC processing and obtain homography matrix based on the specific plane of interest. As a result of the plane segmentation based on fast graph cuts algorithm, this additional operation of plane segmentation can reduce the total time consuming significantly.(2)Build on the foundation of systematic study and research on Homographic Occupancy Constraint (HOC),as well as how to locate multiple objects in crowded scene by using HOC, an improved method which combined the rectangle model and HOC is proposed. Based on the results of the initial target detection, the algorithm selects an optimal reference view for information fusion, thereby obtaining optimum positioning result. Experimental results with detailed qualitative and quantitative analysis are demonstrated in challenging multiview crowded scenes.(3)On the basis of our localization method based on rectangle model, a novel method by establishing a vector table in which the correspondence object ID in other normal views are stored for each object detected in the reference view is adopted to match multiple objects. With the vector table, it is easy to build the correspondence relationship among multiple views.Meanwhile, ghost targets are detected and removed among all the views according to their visibility with the help of the results of target correspondence. Experiments show that this method can eliminate false positive effectively.Finally, the summary of this thesis is made, the shortage of our algorithm is analyzed, and the future direction on the object correspondence is discussed as well based on the work of this study.