The Design Of The Surveillance Videos Database And The Implementation Of The Video Retrieval System

As the raising of security awareness and the development of information technology, more and more surveillance equipment has been installed in public, which leads to the exponential growth in surveillance video data. So it is significant to study retrieval technology among massive amount surveillance videos. On the basis of storage system for massive surveillance videos, this thesis designs and implements a content-based retrieval system.The research of this thesis is organized as follows.Firstly, this thesis proposes an image retrieval method based on ACC-CEDD joint feature, and clarifies the storage method for joint features and the storage format for index files. Meanwhile, it also defines the joint feature, which is the similarity measurement between images. Experiments show that the retrieval method based on joint feature has advantage in precision.Secondly, this thesis has gone through deep research and experiment on video abstraction generation, key frames extraction, and interested targets segmentation, such as pedestrian detection. Then it designs a content-based video retrieval system combining a multi- layer database for surveillance videos and a content-based image retrieval method.Thirdly, this thesis introduces an effective pedestrian image retrieval method. This method uses motion detection and pedestrian detection to extract pedestrian image from surveillance videos. In this way, the obtained pedestrian images are more complete while the whole system’s false detection rate is low. Experiment results suggest that our system can not only search the similar pedestrian image as requested, but also locate the key frame where the pedestrian resides in, as well as provide the detail information about the corresponding video and the key frame. In addition to content-based retrieval, this thesis studies high- level semantic retrieval, semantic annotation system and word segmentation system. The speed for semantic retrieval is fast but the precision is not ideal as expected. So this thesis decides to treat semantic retrieval as an auxiliary search for content-based retrieval to improve the speed and precision.Finally, this thesis designs and realizes a practical retrieval system for surveillance videos on the research of the aforementioned theories. The system, combining Java Web framework development technology and LIRE toolkit, implements user interfaces in front end and the retrieval algor ithms in back end. The system also contains other modules, like user login interface and video management section. Experiments imply that this system is able to retrieve the video and related information in real time with high precision.