The Rise of Terrorist Threats
There has been an exponential increase in the number of surveillance cameras being used for citywide deployment but many recent events have indicated an ineffectiveness surveillance monitoring in pre-empting or preventing such security incidents.
While there has been fewer spectacular terrorist attacks like 9/11, we have definitely seen a shift in the tactics and kind of attacks recently. The Paris attacks were carried out on separate targets almost simultaneously, involved indiscriminate shooting, suicidal attackers and taking of hostages.
The rise in terrorist threats shows the vulnerability associated with public places. They are no longer confined to high value targets like power plants or nuclear facilities. Attacks can now take place easily leaving little time for forewarning of danger even with the proliferation of CCTV.
All these point to the fact that there is now a greater need to discover unusual and deviant behaviour and events as early as possible. The longer the hostage takers are left in control, the higher the resulting death toll. There is a greater need to disperse crowds as quickly as possible and send a response team on the ground to free hostages and reduce casualties.
The Need for Better Analytics
The hard truth is that out of all the thousands of cameras deployed in a city, only a small fraction (less than 10%) are actually looking at perimeters, entrances or exits where an analytic rule can effectively be applied. If we look at this 10% of cameras, we can see that out of the all the possible behaviours or events that can be detected, perhaps another 10% of these events can be detected by applying simple analytic rules. Putting it together means that we are only detecting 1% of all possible events, essentially missing out on the 99.9% of the events that could be of interest to us.
The inherent problem with rule-based analytics is that each rule specified is to detect a specific behaviour (e.g. loitering) and have to make good that claim with 90% accuracy. Much time and effort is then put into configuring this one rule and fine-tuning it to achieve the desired accuracy. If this particular event (i.e. loitering) does not occur, then the rule is basically sitting there doing nothing and produces no value. Worse, if it is inappropriately applied, it produces false alarms instead, adding to frustration of the users.
Revolutionizing City Surveillance –
Doing It Smarter Our product is an unsupervised Machine Learning system that does not require human intervention to automatically discover dominant motion patterns. This also means it does not require a human to specify the rules for event detection. Our Abnormality Detection Algorithm is based on a unique and novel approach. We have adapted it for surveillance videos where multiple motion patterns are occurring simultaneously and it can effectively infer their various patterns and starting times. Since the system is autonomous, it provides the means to automatically analyse hours of video easily. Our proposed system consists of three main components, namely,
a) The automatic abnormality detection (AAD) engine,
b) The rules engine, and
c) The AAD and rules fusion module.
AAD is a completely automatic, unsupervised algorithm to learn frequently occurring activity patterns in the scene that does not make use of any of the existing methods (i.e. Event-based, Rule-based or Object-centric). The functionality of AAD is the automatic detection of abnormal activities by looking for data out of the ordinary. With the set of frequently occurring activity patterns recovered by the unsupervised algorithm, the detection of abnormal activity will correspondingly be automatic. This detection is performed by matching the observed activity against the activity patterns recovered. In contrast to event or object centric methods described in the previous section, this method works automatically without requiring any human input.
- Identify patterns – System is able to automatically identify patterns (i.e. motion, trajectories) of people, vehicles and objects in scenes.
- Discover deviant patterns and behaviours – System is able to find deviations and abnormalities to warn of potential threats (especially those not known beforehand).
- Fast and autonomous – It is able to do it fast and autonomously searching through vast amounts of constantly streaming (or archived) video from thousands of cameras using GPU.
- Not limited by prior human knowledge – Discovery of abnormal events is not limited by prior human knowledge because the system is intelligent enough to self-learn and adapt to different and continuously changing scenarios.
- User feedback – However, the system is able to take in prior human knowledge if necessary. It will speed up the learning for the obvious parts that can be easily accounted for by humans
- Integration with existing systems – We can easily add onto customers’ existing video analytic systems and video management systems so they can increase coverage of their cameras.
- Top 10 search – System is able to list the top 10 most abnormal events based on camera or geographical locations so that users can just focus on these events.