Video analytics, also known as video content analysis (VCA) or intelligent scene analysis, is the use of artificial intelligence to detect activity (changes in the content of a video stream) and then provide verification of the event and create an alarm.
Many algorithms have been produced and they fall into broad categories such as tracking, classification and behavioural. Since the advent of the technology, CHQ consultants have seen algorithms used for alerting on many types of activity or events and in our custodial work we have encountered the problem of prisoners’ associates throwing contraband over perimeter fences.
After being over-hyped at the outset by the sales claims of irresponsible and under-performing vendors, video analytics lost credibility with installers and end-users. But the demonstrable success of a new, robust generation of algorithms has seen the technology’s reach widen from top-tier usage such as the military, transport hubs and critical national infrastructure (CNI) to mid-range applications including general building management, education, logistical processes and non-security usage such as machine vision. Increasingly, the analytical intelligence now resides ‘at the edge’ in the camera rather than in a central core processor.
The global algorithm is the most basic in that it is only looking for simple changes in a field of view. This form of analysis is ideal for creating an instruction to increase a CCTV camera’s recording rate when there is activity in an area that should be sterile.
A basic tracking algorithm will follow an object and determine its direction and speed within a scene. It is not possible for such an algorithm to determine the true speed of an object because it cannot calculate the actual size of the target or where it sits within the depth of field. However, the algorithm can be used to initiate an alert if a fast-moving object crosses a predetermined area in any given direction.
Classification algorithms distinguish between different objects of interest. Currently, best-of-breed algorithms can identify with a high level of accuracy whether a moving object is a person, vehicle or wind-blown debris with further distinctions between, say car or truck.
Once identified, the object can then be subjected to rules such as trip wires, direction, loitering, appearing and disappearing. Thus, a CCTV system used on a railway track to monitor for cable theft would ignore trains as being large and fast-moving but create an alarm when an object with the size and movement pattern of a human entered the field of view. The alarm may be immediate on recognition of this category of object or may be activated when a virtual trip wire is crossed and a sterile zone is entered.
During system installation, behavioural algorithms will monitor a scene over a long period to establish what content and activity is normal and what is atypical and requiring scrutiny with possible generation of an alarm. This process effectively ‘teaches’ the system how to behave. But any external view is likely to involve much ambient movement such as foliage, legitimate traffic flows and wildlife. During commissioning, a behavioural algorithm will ‘learn’ to discount non-threatening, background activity by ‘asking’ questions as to what is normal and what is atypical. The system ‘learns’ what is normal in a scene by employing techniques known as background and foreground subtraction.
In July 2013 the BSIA estimated that there are between 4 and 5.9 million CCTV surveillance cameras in the UK. The Health and Safety Executive is yet to commit itself to estimating the likely effective attention span of a CCTV operative observing a monitor or set of monitors for unusual behaviour though the figure quoted widely in the trade press is 20 minutes with levels of effectiveness deteriorating even during this period.
University research in Johannesburg published in 2011 suggests that 23% of a sample (of whom roughly half were employed full-time in the CCTV sector) lost concentration in the first 30 minutes of observing footage on a video monitor. Multiply these minimal attention spans by the number of known CCTV cameras and the figures for requirement of security guards go off the scale. More precise figures are available from the aviation industry where operatives employed to observe the contents of hand luggage as it is passed through airport x-ray scanners are only allowed to monitor a screen for 10 minutes at a time.
A crucial benefit is that effective video analytics can defend CCTV from suggestions by civil liberty watchdogs that surveillance camera footage is no more than a ‘post-mortem’ tool. Successful case studies in which false (more accurately ‘nuisance’) alarms are minimised through the ability of video analytics software to discount irrelevant activity are now allowing the funding of video analytics from the public purse. At the most simplistic level, it should be noted than the artificial intelligence of video content analysis neither becomes tired nor takes breaks.
In addition to simple reduction of wage bills and freeing up of security guards to do more useful work, other economic cases can be made for the technology. Video analytics can often minimise RAID requirements by allowing a system to record at low resolution until an algorithm detects an anomalous event and increases the recording frame rate. Such techniques can reduce the amount of required storage space by up to 80% on an average system compared to one that is recording continuously.
Increasingly sophisticated modelling of human behaviour is enabling electronic security systems to detect and predict potential activity of interest as it is happening. Certain modes of behaviour can be identified with a high level of accuracy; they include a would-be thief trying successive car doors in a car park, ticket touting at major events and ‘tail-gating’ both through public transport barriers and on toll roads. Claims by one video analytics vendor that there are patterns of behaviour associated with metro or underground suicides – allowing several trains to pass without boarding, posture and proximity to the platform edge, are extremely credible. And yet even the most bullish of video analytics software vendors will concede that a hug can often look like a fight and a handshake can look like a punch. However, the sector is already mature as a new generation of developers react responsibly to the challenge of differentiating objects and significant behaviour from complex backgrounds.
CHQ provides effective technical advice based on the understanding of your threats, the associated hazards and their potential. Working in line with CPNI guidelines, we provide advice and guidance for the security of people and property, critical national infrastructure and the high-security estate.
We provide support in the development and the design of your system by producing schematics, schedules, interaction matrices and configuration tables as well as compliancy documentation and commercial selection.
An effective protection system protects more than just assets; it protects a business.