Video surveillance is becoming automated. It can now track the trajectories of several people at the same time, even when they disappear from camera view for a few moments.

CCTV Surveillance: Facial Recognition Helps Map Movements of Multiple Individuals

Although today the use of closed circuit TV surveillance systems has become widespread, most of the analysis in searching for or tracking people is still being performed manually. Now Alexander Hauptmann, Shoou-I Yu and Yi Yang, three researchers at Carnegie Mellon University, have developed a video surveillance system which can track several people at the same time through the interior of a building. Up to now automated surveillance systems have been tested only in tightly-controlled lab environments, but now the researchers have carried out tests on video archives from at a nursing home which date back to 2005.

Multiple tracking

The US university researchers’ methodology uses several complementary cues to track people’s movements. The person must first be detected and the colour of their clothing is noted, then the trajectory of their movements inside the building is plotted. A fourth cue is facial recognition, which plays a major part in the video tracking. The Carnegie Mellon algorithm can apparently locate people within one metre of their actual position 88% of the time, compared with just 35% and 56% achieved by two of the other leading algorithms used in the surveillance field. However, while facial recognition is a major plus in re-identifying a person moving from one camera to another, it cannot be relied upon throughout the entire tracking process, one of the researchers acknowledging that faces can be recognised in less than 10% of video frames.

Developing the model further

The new algorithm has first and foremost been developed to enable healthcare establishments to monitor the health of their patients. It creates a sort of moving ‘map’ of where a number of identified people are located within the building, the main purpose being to alert healthcare staff to subtle changes in patient activity levels or behaviours that may indicate a change in health status. The researchers say that it could subsequently be adapted for use in airports and public facilities where security is an important factor. This would require further work since, as mentioned above, the study was carried out on video archive dating back to 2005. Future development will concentrate on extending the technique during longer periods of time and enabling real-time monitoring. The researchers are due to present their work at the Computer Vision and Pattern Recognition Conference in Portland, Oregon on 27 June.


By Guillaume Parodi