Go to most cities and you’ll find tens or hundreds of thousands of surveillance cameras helping the authorities to keep the peace. The number of visual feeds – combined with their typical around-the-clock operation – is usually too high for government agencies to monitor at all times.
Now, Professor Wang Gang and his researchers at the Rapid-Rich Object Search (ROSE) Lab at School of Electrical and Eleectronic Engineering have developed revolutionary computer algorithms that can give law enforcement agencies a powerful helping hand. These algorithms can sift through cameras’ views, detect people even if they are partially hidden, track them across cameras and flag suspicious activity.
The researchers used deep machine learning techniques, where machines learn from massive datasets and past experience to better sort through incoming information and extract useful information from raw data. Once trained, the software could monitor camera feeds on their own and alert security officers when needed (a surveillance scenario), or to search through video records to retrieve images of persons or objects of interest (a forensic scenario).
The software detects all of the people in a scene, even partly obscured ones, assigns a unique identifier to each person and tracks their movements. Even if there are 50 people in the camera’s view, the program can follow each one – crucial for monitoring popular places such as tourist spots.
When a person steps out of the scene – say, to go into a shop – and reemerges later, the algorithms will recognise him or her and re-assign the same identifier, rather than a new one. They can also re-identify the person across multiple cameras and despite lighting changes that might alter their appearance.
Professor Wang Gang from School of Electrical and Electronic Engineering said: “Let’s say a murder is committed, you see the culprit on camera 1, and you want nearby cameras to look for this person. They might be looking onto places in shadow or bright sun, which will cause the person’s clothes to look different. The algorithms can still identify and track the person across cameras.”
The algorithms can also recognise what people are doing, say if two people are fighting or just shaking hands. This will be useful for security officers. The software uses two different types of cameras’ information – typical RGB cameras and depth-sensing ones – to recognize an action or differentiate between similar ones.
The software can also be used to count crowds, analyse their behaviour and other purposes. Professor Wang said: “In future, intelligent systems will complement human effort on a larger scale. We believe this advancement is key to fulfilling the vision of a Smart Nation.”
By Prof Wang Gang, School of EEE
Click here to find out more.