Abstract

Modern surveillance systems allow high quality video data to be stored in large quantities. Users can search through the archive to retrieve content of interest. However, this search is limited and time-consuming, unless the user can be helped to find relevant content. This can be via a response to user-centric queries, or more simply by generic highlighting of content that is considered of interest (e.g. anomalous). Thus, just as internet users will use a search engine to find, display and select appropriate web content, video surveillance users could find utility in a corresponding search box for that domain. Furthermore, the video data search engine will have to process or 'index' the video data, to provide a timely response to the users' queries. The research programme will address at least one of the following specific issues: - The search and retrieval process is limited by the sophistication and accuracy of the features extracted for indexing. This will depend on the quality of the data (including the quality of view and degree of crowdedness), and the complexity and fragility of the information extraction process. - Efficient video indexing is a necessity. The standard case is to be able to perform indexing in real time, to allow the current 'live' streams to be indexed for subsequent retrieval. There are issues to be explored here, such as the preliminary analysis of video streams in the compressed domain (since this would be more efficient). Furthermore, for archives that are not currently indexed, then technology to perform preliminary indexing in 'super-real time' (e.g. indexing 3 weeks of data overnight) may be an important feature in certain post-event scenarios. - Another type of data mining does not require user queries: this is the search for 'anomalies' – events (or data) that is unusual in some important respect. - A particular forensic investigation may encompass large amounts of video from multiple and heterogeneous sources. One issue is the efficient navigation, viewing, and cross reference of the various sources. To this end, it is useful to determine the degree of overlap, the degree of synchronicity, and the variations in quality that will exist between them. Most importantly, the video data mining project will benefit from an investigation into the most appropriate user interfaces to facilitate a forensic search. Possible focus areas: - Investation into novel space-time representations to show activity and indicate times and places where activity may be relevant to user query and to allow integration and evaluation of video data from heterogeneous sources - Investigation into selection of features that can be processed in real time and/or "super-real time" to allow for efficient indexing of the archive. NB processing in the compressed domain. - Investigation into design, selection and test of features that would be useful to perform data mining on urban traffic scenes, e.g. locating vehicles of interest or searching for particular types of behaviour (I think vehicles is useful subject for this data mining, as they are rigid, fairly distinctive and there is a reasonable upper bound on how crowded they can get!) - Investigation into video mining techniques for detection of anomalies.