Research On Video Synopsis Technology In Multi-Camera Surrounding Environment

With the widespread use of web cameras and surveillance cameras,mass real-time video streams are being generated all the time,which makes the retrieval of useful targets more and more challenging.To address this problem,video synopsis technology is proposed,especially the object-based video synopsis method.The key idea is to first extract the tubes of all moving objects appeared in different time instants and then to re-layout these tubes in a shorter duration.In this way,we can quickly browse all key targets and their trajectories.However,most of existing works are designed for synopsis of video from a single camera,which cannot work well on the multi-camera monitoring network.In this thesis,we attempt to study the video synopsis technology in multi-camera surrounding environment.The main contributions can be summarized as follows.1.Propose a novel video synopsis technology for multi-camera monitoring networks,named surrounding multi-camera video synopsis.Different from traditional single-camera based video synopsis methods,the proposed scheme uses probabilistic occupancy map and multiple object tracking method to track the activities of objects in multi-camera monitoring areas effectively.In this way,a complete object activity tube can be extracted for each object by using multiple videos from different cameras.Then,a new energy function is proposed to re-layout the extracted tubes and a shorter video is then generated.Experimental results show that the proposed method indeed summarizes the long video stream effectively.2.A new multi-camera video synopsis system is designed and implemented so as to intuitively show the effectiveness of the proposed method.To provide a more friendly interface and show the results more intuitively,we present objects in the form of a 3D rectangular surface in the simulated 3D pseudo-scene space.The shift of object is manifested by moving the position of the rectangular surface on the simulated ground plane.And then,the physical activities of the objects are expressed by rendering the images of object's activity on the rectangular surface.In this way,the users can observe the objects from different positions and views and avoids to be occluded.