- U.S. funding research into AI that can predict how people will behave
- Software recognises activities and predicts what might happen next
- Intended for use in both military and civilian contexts
By Damien Gayle
PUBLISHED: 05:59 EST, 23 November 2012 | UPDATED: 06:09 EST, 23 November 2012
An artificial intelligence system that connects to surveillance cameras to predict when people are about to commit a crime is under development, funded by the U.S. military.
The software, dubbed Mind’s Eye, recognises human activities seen on CCTV and uses algorithms to predict what the targets might do next – then notify the authorities.
The technology has echoes of the Hollywood film Minority Report, where people are punished for crimes they are predicted to commit, rather than after committing a crime.
Scientists from Carnegie Mellon University in Pittsburgh, Pennsylvania, have presented a paper demonstrating how such so-called ‘activity forecasting’ would work.
Their study, funded by the U.S. Army Research Laboratory, focuses on the ‘automatic detection of anomalous and threatening behaviour’ by simulating the ways humans filter and generalise information from the senses.
The system works using a high-level artificial intelligence infrastructure the researchers call a ‘cognitive engine’ that can learn to link relevant signals with background knowledge and tie it together.
The signals the AI can recognise – characterised by verbs including ‘walk’, ‘run’, ‘carry’, ‘pick-up’, ‘haul’, ‘follow’, and ‘chase’, among others – cover basic action types which are then set in context to see whether they constitute suspicious behaviour.
The device is expected to be used at airports, bus and train stations, as well as in military contexts where differentiating between suspicious and non-suspicious behaviour is important, like when trying to differentiate between civilians and militants in places like Afghanistan.
KILLER ROBOTS ‘IN 20 YEARS’
Fully autonomous robots that decide for themselves when to kill could be developed within 20 to 30 years, or ‘even sooner’, a report has warned.
Militaries across the world are said to be ‘very excited’ about machines that could deployed alone in battle, sparing human troops from dangerous situations.
The U.S. is leading development in such ‘killer robots’, notably unmanned drones often used to attack suspected militants in Pakistan, Yemen and elsewhere.
While these are remotely controlled by human operators, autonomous robots would be left to make highly nuanced decisions on their own, the most fraught being distinguishing between civilians and combatants in a war zone.
The warnings come from a report by Human Rights Watch, which insists that such Terminator-style robots are banned before governments start deploying them.
The paper describes an example in a civilian context where it might flag up a person dragging a heavy bag then leaving it abandoned for more than a few minutes.
According to Phys.org, this automated approach to surveillance could one day tempt authorities to replace humans with computers as CCTV camera operators.
Human operators are expensive to maintain and also fallible: distracted or drowsy operators pose a risk to safety if they miss danger signs, while ever-vigilant computers, once installed, are cheaper to employ.
Also, cameras that do nothing but record what they are seeing can only provide information after a crime has occurred, while the latest research hopes to use unmanned cameras to prevent crimes or dangerous behaviour happening at all.
Alessandro Oltramari, a postdoctoral researcher and Christian Lebiere, both from the Department of Psychology at Carnegie Mellon, a just one team of a number working on automatic video surveillance for the U.S. military.