Getting smarter about video analytics


Singapore, November 29, 2016


In this month’s Q&A, we spoke with Mr Sagin Hsu and Mr Ching Yin Sing to find out more about the state of video analytics technology, how machine learning can improve its accuracy and what organisations need to know before implementing video analytics systems. The technology, while useful for certain situations, still had some ways to go before it can accurately perform complex tasks such as determining human behaviour.
 
Getting smarter about video analytics
 
Q: More public safety agencies and facility managers are turning to video analytics, which can automatically detect suspicious activities and vehicle number plates to keep our public spaces, buildings and roads safe. How accurate are video analytics solutions today?
A: For a start, video analytics is the use of software to analyse video footage of events that are of interest to organisations like public safety agencies. These can range from simple quantitative tasks like counting the number of people at a location to more complex tasks such as recognising a face from a crowd and analysing behaviours. Today, the accuracy of video analytics solutions tends towards quantitative tasks and gets less accurate when applied to more complex and qualitative tasks. That’s because in analysing, say, human behaviour, it’s hard to ascertain if a group of people are indeed fighting, or just fooling around. .
 
Q: So does that mean there’s still some ways to go before video analytics technology becomes more mature?
A: Today, video analytics is most apt in handling quantitative tasks, followed by “middle of the road” applications like facial recognition, the accuracy of which can be dependent on environmental conditions such as lighting. So, in a well-lit controlled environment with people looking at the camera, you can accurately detect faces, but less so in dim lighting. As for complex tasks like detecting suspicious behaviour, more work and research needs to be done
 
Q: How can machine learning improve the accuracy of video analytics solutions?
A: Machine learning techniques have been increasingly applied in video analytics applications. This has greatly improved the accuracy of facial recognition and licence plate recognition. Machine learning allows the computer to learn from data presented to it in an iterative manner without the need to explicitly programming by a human on what to look out for. 
 
Q: But wouldn’t that require heaps of data, especially for training video analytics applications to detect human behaviour?
A: For machine learning to occur, there is a need to provide sufficient data that enables a machine to tell and learn, such as in image recognition. If you don’t have enough video data, it would be difficult to train the machine.  It should be noted that while an algorithm may detect a certain behaviour, it is not possible to elicit the motivation behind the behaviour. 
 
Q: What’s the thinking behind the data fusion concept?
A: Data fusion refers to the taking in and manipulating different type of data regarding an event or situation so as to provide more accurate information. For example, combining video footage of smoke with alarms from smoke detectors, together with social media feeds of fire reported at the same location, can provide confirmation of a fire incident.
 
Q: What should organisations consider when implementing video analytics?
A: Organisations need to be clear about the use cases that they are interested in and need to be realistic in their expectations of the technology. If they intend to use video analytics to solve a highly qualitative problem then it is likely that video analytics may not be suitable.  Regardless, there are five broad steps that can be taken when implementing video analytics. These are:
1) Select the most suitable video analytics software for the use case being considered.
2) Select and if needed customise the required cameras and sensors.
3) Tune, configure and test the system. This step cannot be over emphasised as video analytics is dependent on tuning.
4) Fuse in data from other sensors to provide other useful information.
5) If needed, develop specialised algorithms to integrate the system.