Have you seen this person? The Chinese People鈥檚 Liberation Army (PLA) is building artificial intelligence that can answer such a question.
Despite privacy concerns over CCTV, the task of finding specific individuals in captured footage, or pinpointing people caught on camera in different places, is a tough one. There鈥檚 no telling how differently someone may appear, or behave, from one听camera to another.
However, the PLA has designed听a system that听seems to be able to听track people across different sources of CCTV footage听with around 90 per cent accuracy.
Advertisement
Their tool, EnsembleNet, was trained on CCTV clips from public places such as . In total, the footage features more than 2,000 people.
The tool works by breaking these images down to discernible body shapes at various levels of detail. These听 kinds of analyses allow for figures appearing elsewhere on camera to be compared in听different ways.
The head and legs, could match even when the torso doesn鈥檛 seem to, for instance. The authors write that this means re-identification can be done 鈥渞egardless of the pose or gait鈥� of the pedestrian.
As an example, the system was able to re-identify a man who was wearing a white card on his front even when he constantly had his back to the camera in later footage. A distinguishing mark like that could have thrown lesser systems off, but the authors say this shows how EnsembleNet can pick people out despite subtle differences in appearance.
Such technology is a focus for many researchers in China, where the tracking of people and vehicles is an increasingly automated pursuit. In Shenzhen, face recognition technology is used to spot jaywalkers and automatically send them fines.
鈥淭he seemingly state-of-the-art nature of this particular project is another indication that Chinese military has the capability to achieve notable progress,鈥� says at the Center for New American Security.
She points out that Chinese authorities have been keen to ensure that various large datasets be made available for computer vision research projects 鈥� including surveillance.
Reference:听
