Article Preview
Top1. Introduction
CCTV systems are playing an increasingly important role in remote surveillance of people, property and public and private sites. Their first appearances were in the 1950s. However, surveillance was really developed from the 1970s using closed route TV (CCTV) systems, mainly in the United Kingdom. The implementation of video surveillance intensified throughout the 1990s. Since the 2001 attacks in the United States and 2005 in London, the number of surveillance systems installed has increased. Thus, a considerable number of cameras have been widely deployed in public sitting room, including transport infrastructure (metro stations, airports, ...), parking lots, banks, shopping centers, roads and industrial sites as a instrument for crime reduction and danger management (Chebi 2016; Chebi 2018;Gouaillier 2009).
In the case of video surveillance applications, the objective is most often to be able to detect events of interest such as, for example, the detection of accidents, thefts, or the attendance of illegal persons. The classic exercise of these recorded scenes is also to have a human operator monitoring the screens directly connected to the sensor via a high-speed network, or to use the videos a posteriori to, for example, find the author of a theft. To allow more efficient use of the available video streams, it is necessary to design methods capable of automatically detecting events of interest and to have an effective compression method (H.264 type for example) when the available bit rate does not not allow raw videos to be transmitted in real time. So it's about recording, compressing, transmitting, decompressing and analyzing video in real time. However, in many cases, only a very small proportion of the information transmitted is really useful, i.e. the areas in certain portions of the video showing an event of interest. To reduce the flow of video data before being transmitted, one solution consists in analyzing the video scene at the level of the sensor, and in transmitting after any compression only the useful information (regions of interest). As the analysis algorithms can be complex, such a solution may require significant computing resources at the sensor level, which is not often possible. To remedy this problem, our solution consists in detecting, by means of a pre-analysis phase, at the level of the camera the zones likely to contain useful information, then in analyzing these zones to eliminate the information which is a priori useless. Finally, only these simplified zones will be compressed to be analyzed more finely at the reception level. To be effective, the pre-analysis phase at the sensor level must be relatively light in terms of computational load, and not destroy the information useful for event detection. Consequently, the usual scheme of video surveillance systems is improved, as shown in Figure 1, to become: record, pre-analyze, compress, transmit, decompress and analyze the scene. Adding the pre-analysis phase to the usual process requires additional calculations at the sensor level. On return, it accelerates the transmission and analysis of data. In typical hardware architecture, video streams are stored on a central server and then distributed for analysis. Since in our proposed approach the pre-analysis must be performed on the sensor side, this architecture is no longer adequate. A distributed architecture where video storage and pre-analysis are performed at the sensor level is suitable (Senior 2009). Information is only transmitted when requested by the user.
Figure 1. Proposed improved architecture of video surveillance systems