Keeping it Real-time : advances in video analytics for CCTV
This year, it is predicted that almost a million minutes of video content will circulate the internet every single second. That is a lot of data – and a lot of potential analysis.
The dramatic growth in the use of this medium is driving the demand for more efficient ways of analysing video data, particularly for CCTVs in the security industry.
“There is a growing need for video analytics,” explains Dr Victor Sanchez, from the University of Warwick’s Department of Computer Science.
Current analytics often sees a trade-off between quality and reliability on the one hand and on the other hand, computational power and storage. So, at the moment, analytics are not widely used in train stations, airports, highways, office blocks, housing developments, retail applications and distribution centres.
“We wanted to investigate a solution that would help anyone looking for efficient real-time analytics. We wanted to design a tool which would provide a ‘second set of eyes’ – a back-up for anyone with the difficult task of constantly scanning multiple, complex screens streaming security footage from a number of different positions.”
Machine learning
Currently, the application of deep learning analytics – where computers can predict or detect an event, based on the rules ‘learned’ from earlier data – are providing a number of solutions. These methods are either based on Convolutional Neural Networks (CNNs) or hand-crafted feature descriptors.
“CNNs have demonstrated outstanding performance for video analysis tasks,” continues Dr Sanchez. “But these algorithms take a relatively long time to get up and running, because there are many layers to the data. It can take days or even months to train a descriptor for something like action recognition.
“Those methods in existence which do require shorter training and processing times, for example tools which use feature descriptors, are highly dimensional in terms of the extracted features and therefore require vast storage capacities and computational resource. This makes them expensive and demanding systems to run.”
New video analytics
The research team at the University of Warwick set to work investigating a new method of feature descriptor for video analytics.
“We looked at a method which encodes the motion information of a Spatio Temporal support region into a low-dimensional Binary string (STB),” explains Dr Sanchez. “The encoded information for the process is obtained from two motion sources – the optical flow and the temporal gradients.
“Using two motion sources provides rich motion information by considering the pixel intensity changes to create a new data space that disregards the background. The tools can therefore quickly detect differences in pixel intensity that are likely to be an abnormal change in a given picture and alert to a possible security breach.”
Promising results
“We are delighted with the results our model is producing so far,” says Dr Sanchez. “It can be used across a wide range of uses including classification, activity recognition, object recognition, video surveillance, monitoring and video anomaly detection, and it is suitable for use in real-time situations. The biggest advance, however, has been successfully reducing the demand on training time to hours, computational power, storage and memory. It means the tool could now run on something like a mobile phone or a CCTV camera itself – no longer needing a powerful computer.”
The team’s findings show huge promise for the security industry.
Dr Sanchez continues: “There are a number of big advantages for a company wanting to employ this kind of solution. 24-hour surveillance is a demanding job that will still need a human eye, but using an automatic real-time monitoring system in the background would give added back-up.
“In addition, a single operator may have the responsibility of watching many screens, streaming many views simultaneously, but the human eye cannot pay attention to every single camera and so a real-time monitoring solution would help to do the job more efficiently. Plus, the low computation power requirement of the tool would allow it to be run on a mobile phone, making it portable and freeing up security personnel to move around a given area to make physical checks.”
Dr Victor Sanchez
Researcher from the Computer Science department at the University of Warwick.
This invention is now the subject of a patent application PCT GB2018/058103, and the researchers are working with Warwick Ventures to explore licensing opportunities for this new solution.
For more information visit: https://warwick.ac.uk/siplab/research/
