Video surveillance systems have moved on significantly in recent years, both in public applications and on industrial sites.
Continuous innovations in terms of image quality and resolution (with the move to high definition digital video) and connectivity (with IOT), are further empowered by the ease of storage in the Cloud.
Today the extensive development of computer vision applications, with their huge potential for detection, recognition and prediction, enables the video surveillance market to take another leap forward.
How deep learning makes video surveillance more intelligent
The video surveillance sector offers numerous applications for artificial intelligence. Machine learning and deep learning algorithms convert video surveillance systems into authentic visual intelligence ecosystems capable of recognising and analysing events and behaviour with an accuracy and rapidity that were impossible hitherto, and unattainable by the human eye alone.
Where the human eye and brain are shackled by their physical and physiological limitations (visual capacities, tiredness, loss of concentration, etc.), artificial intelligence is limited only by the quality of its training.
Indeed, training data plays a key role: it enables you to educate your algorithms in order to fully exploit their possibilities to provide the additional level of intelligence, automation and prediction to your video surveillance system.
Better targeting of relevant elements
IC Realtime has developed Ella, a visual search platform which functions like a search engine. It enables the result of a request, expressed in natural language such as "red truck", to be found almost instantly among several hours of video streams. In practical terms this means that it allows you to find what you are looking for more easily and more rapidly without needing to view hours of irrelevant images.
An ethical issue
Today these models permit the detection of movements and the analysis and recognition of behaviour suggestive of an event that has the potential to be a threat to personal safety. However, and even if the Chinese start-up SenseTime is currently valued at $4.5 billion, facial recognition technologies raise serious ethical issues.
Artificial intelligence supporting more proactive security systems
Surveillance automation is moving towards a predictive approach through what is known as "Intelligent Motion Detection". This aims at identifying problems before they happen through real time detection of known subjects identified as suspects.
The applications are unlimited: the prevention of school bullying or unruly crowd behaviour during meetings and demonstrations, road accidents (e.g. by comparing dangerous driver behaviour with traffic and weather data), etc.
A team of researchers trained a deep learning algorithm using the Aerial Violent Individual dataset, comprising 2,000 labelled images of drones presenting groups of persons, some of which displayed violent behaviour. This led to the development of Eye in the Sky, a system of surveillance by drones capable of real time detection of violent individuals in a crowd.
Today the application of deep learning to security video surveillance is booming. Companies like Viisights, Calipsa or Cogisen, for example, provide intelligent video monitoring solutions based on machine learning models.
The integration of models like these in the latest generation surveillance cameras allow the operators to monitor situations in real time and identify problems even before the happen.
For example, the Japanese railway corporation West Japan Railway uses an artificial intelligence module that detects intoxicated users from their behaviour and posture so as to prevent accidents involving people falling onto the tracks - of which they are the main victims.
Surveillance of public areas
Meanwhile, Movidius and Hikvision have cooperated to produce a range of cameras connected to a deep neural network for automatic detection of suspicious behaviour or objects that a human operator might not have identified. These solutions offer the advantage of being able to recognise "abnormal" behaviour or parameters in a store, a bank, a public area, or an industrial complex, etc.
Purchasing behaviour analysis
It also has an interesting application in retailing, and the analysis of behaviour for marketing purposes. The traffic in store aisles, together with consumers' gestures and choices in each section are a mine of information for store holders and the mass retail market.
While it doesn’t – yet – replace the human element, machine learning does ensure increased reactivity and accuracy in monitoring operations and, through its processing power and continuous learning abilities, brings an additional degree of intelligence and autonomy to video systems. Did you say Big Brother?