Video Retrieval Research For Public Safety

With the fast development of multimedia and network technology,as well as the emergence of various video compression and mass storage technologies,the acquisition,dissemination and storage of video information have been more and more simple.At the same time,the security issue has attracted attention all along.As a key part of the security system,the efficient and rapid retrieval of the pedestrians inside the surveillance videos has become an attracted topic in the field of content-based video retrieval.But due to the poor resolution and light transformation of the surveillance videos,and the appearances of pedestrians are susceptible to the wear,perspective and other factors,this technology still faces a huge challenge.In this paper,the technologies of pedestrian detection,feature extraction and similarity discrimination are studied.And a pedestrian retrieval system based on pedestrian feature and metric learning is designed and implemented.Firstly,by comparing several classical background subtraction and statistic learning based methods in the existing pedestrian detection,a new method based on statistical learning and background subtraction is proposed to improve the pedestrian detection effect.Afterwards,the LOMO,main color,shape and face features are extracted.At the same time,in order to reduce the repetitive retrieval work,a key frame extraction method based on the moving object area and the proportion of the face is proposed.The face parts of the key frames are dealt with the interpolation processing.Then,the pedestrian re-recognition based on metric learning is studied,which realizes the study of similarity distance and the dimensionality reduction of features.By carrying out experiments on public datasets,the experimental curve of CMC(Cumulative Match Characteristic)indicates that 90%of the top twenty search results are correctly identified.Finally,a web based client-server user retrieval system is designed and implemented.The system provides two functions:the fuzzy and image query.The fuzzy query uses the main color features combined with the video time as inputs,and the image query uses a single pedestrian picture as an input.Experiments show that the overall system recognition rate can reach 80%.