Hong Kong researchers develop 3D tool for surveillance footage
HKU team refines system that cuts time police need to sift through security-camera images
Watch: How HKU's 3D surveillance reconstruction system works
Researchers are refining a system that slashes the time investigators need to sift through hours of security-camera images from crime scenes.
A University of Hong Kong team is working on a way to merge footage from different angles into a three-dimensional reconstruction of the scene.
Police officers - one serving and another retired - said the technology was not new. The serving officer said he was aware of similar software being used by the FBI in the United States.
While it now can take several hours to process a single hour of footage, the Hong Kong researchers say they have got the time down to 90 to 120 minutes, depending on the complexity of the scene, and hope to take it lower.
The project - which began in June 2012 with most of its HK$5.22 million funding coming from the government's Innovation and Technology Fund - lets users closely examine images from all available viewing angles.
Once one or more suspects have been identified, the system will sieve through all available footage and zoom in on the segments in which suspects are captured, Professor Yu Yizhou, from HKU's department of computer science, explained.
The system tracks the suspects identified based on features such as the colour and texture of skin, clothes or personal belongings, as well as the way they walk.
With four cameras filming the lobby of a HKU building, the accuracy rate of the system was about 90 per cent.
Accuracy is slightly lower - about 80 per cent - if the location is crowded most of the time, though more cameras could improve the accuracy.
Yu said the system would primarily be used to reduce the possible number of suspects and save investigators' time.
Feedback was positive last month at a briefing on the new system for police, customs officers and the Airport Authority.
At present, painstaking study of footage is needed to identify suspects manually.
Investigators are said to have spent hundreds of hours examining footage to identify the possible assailants of former Ming Pao chief editor Kevin Lau Chun-to, who was ambushed and chopped from behind in Sai Wan Ho on February 26.
The research team behind the system has yet to test it using footage below the standard definition quality, or to put it to use on images recorded outdoors.
A trial version will be available in a year but "several million" dollars is still needed before it can be put on the market.
Other aspects that help identify a person, such as body shape and posture, will also be added to improve accuracy.