Effective CCTV Target Tracking
Automated tracking of targets such as individuals and vehicles has been a goal of those providing security solutions for many years. While manufacturers have provided various options that claim to deliver automatic target tracking, such systems are still a rarity in the field. Benchmark considers the role of CCTV tracking and target recognition systems to see whether installers and integrators should be making more of the technology.
There are certain jobs that even those who love technology have to admit are better done by a human. In the past, one example was CCTV tracking. Indeed, despite the many flaws that are inherent in human nature, plus the reliability failings that are well documented when man takes the place of a machine, humans are currently superior in terms of object recognition and classification.
A machine could, for example, recognise a red car. It could even apply shape- and size-based criteria to that recognition, for example allowing it to correctly recognise that a red SUV was not a red sports car. However, would it be able to spot that a red Peugeot 308 is not a red Ford Focus? It might, given time nd complex algorithms, but that’s of little use in a security application.
A system operator could probably tell the difference. Even if they were not conversant in car identification techniques, other details – damage to the vehicle, specific trim, number of occupants, etc. – would help them to reasonably recognise that two red vehicles in the same area were not one and the same.
When it comes to tracking targets in a viewed scene, the role traditionally falls to an operator. Across the world there are control rooms with personnel taking control of PTZ cameras and following targets – whether individuals, vehicles or other moving objects – for a whole host of reasons. Whether the goal is security, safety, asset management or logistical verification, the task relies upon the attention and recognition capabilities of the operator.
The understanding of a site by an individual becomes even more critical when the tracked target moves around the site. In order to ensure continuity of tracking, the operator may have to switch between different cameras to keep the target under surveillance.
If a ‘blind’ area exists in a site, then the operator needs to know which cameras on the system could potentially pick up the target next. This would allow them to check these to re-establish surveillance as and when the target once more was captured by the surveillance cameras.
As the potential offered by video analytics increases, so automatic tracking has increasingly been offered as a feature of PTZ-enabled cameras. Whilst many of the analytic offerings have been embraced by installers and integrators, auto-tracking isn’t as frequently used in either live viewing or recorded analysis applications.
It is fair to say that auto-tracking does present a number of challenges. The ‘hand-over’ between different cameras in a system is one that manufacturers have worked on, but certain assumptions must be made if it is to be successful.
Firstly, the surveillance on a site must offer a degree of continuity. Whilst it is ‘good practice’ to ensure that the coverage of cameras covers all areas and even has a degree of overlap at the edges of the field of view, such an approach simply doesn’t translate into reality.
End users are ultimately the people who decide the degree of security that is implemented at their sites. Installers and integrators can make recommendations, as can specifiers and consultants, but the final decision rests with the customer. Their decision will (hopefully) be based upon a risk analysis. This will typically be based upon information that the end user is the best position to analyse. If they decide that there is a minimal risk in certain parts of the site, then there simply will not be continuity of surveillance.
This lack of continuity may be further compounded if there is a risk of staff involvement in nefarious activities. This may be direct involvement or collusion.
In such applications, any hand-over between cameras is going to difficult, and immediately raises issues for automatic target tracking implementations. This can be highlighted by an old scam from the warehousing sector which was used to trick human operators.
Where PTZ cameras did not have an overlap, an overall-wearing member of staff – taking care to hide their face from camera – would collect goods from a secure area in a box or bag. When in the blind area between the cameras, they would stop and a similar looking employee would carry on towards the site exit. The operator would track them and inform personnel at the gate to carry out a search, which revealed a box or bag of rubbish. The real target took advantage of the lack of security and slipped away.
Any automated tracking hand-over that cannot establish continuity leaves itself open to questions about its credibility if footage is used as evidence. Many cases that go before the courts and fail do so because of a lack of continuity.
Going back to the scenario with two different but similar looking red cars, there are potential issues that need to be considered. For example, if Car A – the target vehicle – is heading out of a site and Car B – an innocent but very similar vehicle – is entering the same site. If they cross paths in a blind area between two camera fields of view, what will the outcome be? Will the first camera’s analytics see Car B as Car A having turned around, or will it hand-over the tracking?
This concern might seem to be based on a huge coincidence, but few analytics programmes use colour as a filter. Also, vehicle sizes will either be manually set as a maximum and minimum size to allow for all variations, or a self-learning system will ‘approximate’ a vehicle size. In short, as the likelihood of a problem increases, so does the challenge for any automatic tracking solution using video metadata as its core intelligence.
Conversations with several installers and integrators reveals that many do not consider automatic tracking because of bad experiences. Erratic handovers and switched targets are the two main issues. With target switching, the analytics might be locked on to an individual, but as they pass another person – often behind the second individual – the analytics ‘switches’ the tracking to the innocent person.
Whilst high-end analytics do allow more in-depth filtering top include colour, shape and size correlations, speed, distance, etc., the reality remains that for the majority of applications based upon video analytics, automatic tracking can be hit-and-miss.
A different approach comes through the use of detection that can attribute exact locations to targets. One example of this is laser-based detectors, which identify the location of a trigger object using X-Y co-ordinates.
A site plan can be created using GPS mapping, which can cover both internal and external areas. In effect, every pixel can be attributed a GPS co-ordinate. The laser detector’s X-Y co-ordinates in relation to the tracked motion are translated to this map, which produces accurate information about the exact location of any target.
The data can then be used to trigger and drive absolute-positioning PTZ cameras. These can track the intruders through the site, and if a group of people are detected and then split up, the software can take a central point and zoom the cameras accordingly to ensure an operator can see what is happening. Additionally, the software allows the system to ‘hand over’ to other cameras and can allow for video blind spots so long as the coverage of the laser detectors is complete.
The one down-side of this approach is that tracking cannot consider any ‘visual’ data about the targets. To a laser detector, a target is simply an object located at a co-ordinate, and no amount of filtering can be established for tracking purposes. It relies upon an individual viewing footage to make any reasoned decisions.
A sterile approach
Whether you are considering automatic target tracking using video analytics or detector-based co-ordinates, the most important issue is to understand the potential problems and false activations that might occur.
As with many analytics- or detector-driven solutions, auto-tracking works best – in both live and recorded video applications – if deployed in a sterile zone.
A sterile zone consists of some form of obvious demarcation between an area that people are permitted to be in, and an area which they clearly excluded from. The idea of a sterile zone is not to physically restrict access to an area – although some sterile zones do just that – but to ensure that anyone entering the zone has done so with intent rather than accidentally. Typical approaches include fencing or other barriers, signage or audio-visual devices.
By creating a sterile zone, this establishes that any activity in the area represents a violation, and this fact alone simplifies the creation of video analytics rules of all types.
It is the simplicity of the sterile zone which makes it so effective when compared to scenarios with a greater number of variables. If someone or something is in the sterile zone, it is an alarm condition.
Of course, if a sterile zone is used, tracking as an investigatory tool is somewhat unnecessary. A violation is just that, and an alarm event can be generated and acted upon.
Automatic tracking is a specialised tool, and if a system has been designed to support the function it can work. However, if the vast majority of sites there are more cost-effective options which can deliver results. The technology is still being developed, but the huge number of variables at most sites limit its appeal, especially given the enhanced performance from other IVA rules.