During a recent webinar with the International Institute or Risk and Safety Management (IIRSM) about plant – pedestrian safety I was asked this question:
“What differentiates you from your competitors, for example the emergence of AI and AI Systems and other technologies that are available.”
It’s not uncommon to be asked this type of question having been involved in mobile plant – pedestrian safety for over 12 years now. But I have to be honest, on this occasion my answer was not as well defined as I would like. I know all about the prominence of “Artificial Intelligence (AI)” cameras (we’re using that phrase loosely for now), but what else did I know?
A little more than I thought apparently.
More than a decade ago I was involved in introducing a new mobile security CCTV camera into the UK market. At that time, it was impressive. Someone entered the detection field of the camera, a little box appeared around them and then tracked their movements while they remained in the field of detection. Sound familiar?
In 2012 we called this type of technology Video Contents Analysis (VCA).
So, I did some more digging (and even asked AI in the form of Chat GPT for some help!)
Let’s start with the true definition of AI, if these systems really are AI.
Stanford’s AI Index 2023 says:
“True AI (Artificial Intelligence) implies a level of autonomous learning and decision-making that goes beyond mere pattern recognition .”
So, do the current systems out on the market for pedestrian detection AI have a level of autonomous learning? I couldn’t find any evidence of this. In fact, what I found was remarkably like the technology we used all those years ago.
If these pedestrian detection systems almost certainly do not meet the definition of true AI, how do they work?
AI pedestrian detection cameras analyse video footage to identify pedestrians within a given frame. They do use machine learning algorithms trained on a database of images containing various images of pedestrians, lighting conditions, and environments. By comparing the live video to this database, the system attempts to recognise and classify objects as pedestrians.
These systems are employing advanced VCA techniques. They match what they “see” (live footage) against a static database of known images, relying on predefined patterns and characteristics.
Limitations of Video Content Analysis (VCA) or human form recognition
While video content analysis is powerful, it does have limitations, particularly when the database of images it relies upon is not comprehensive and the environment is complex and changing.
- Limited Database
The usefulness of these systems relies on the size and diversity of their database. If the database lacks sufficient variety in terms of pedestrian appearances, clothing, and behaviours, the system may fail to accurately detect pedestrians in real-time.
For instance, individuals with unusual clothing, carrying large objects, or moving in unusual ways might not be recognised correctly.
Clearly these databases can be updated, however it raises the question; How are they updated once a system is installed and out in the field?
- False Positives and Negatives
A false positive is when the system incorrectly identifies a non-pedestrian object as a pedestrian. Conversely, a false negative occurs when the system fails to detect an actual pedestrian.
These errors generally occur due to environmental factors like shadows, reflections (think about hi-vis clothing and cones used in most environments), and poor and changing lighting conditions. These can confuse the algorithm.
- The operational environment and application
Pedestrian detection cameras may struggle in complex environments, particularly as they have more environmental factors to deal with. Whilst it’s simple to add a something like a traffic cone to the database, so these are not detected, non-standard objects such as irregular piles of waste on a waste handling site or “shapes” which are constantly changing are harder to add. Varying lighting conditions present a significant challenge. Such as going from inside to outside.
- Static vs. Dynamic Learning
Unlike true AI, which can adapt and learn from new data over time, many pedestrian detection systems are relatively static. Once trained on a specific dataset, their ability to improve and adapt without further manual updates is limited. This means that new pedestrian behaviours or changes in environmental conditions may not be effectively handled.
Conclusion
This is not a piece to say Pedestrian Detection Cameras are either good or bad. Having seen the result and impact of plant – pedestrian collisions anything which can effectively and repeatedly detect pedestrians and prevent collisions is a positive.
I wanted to understand what these systems really are. And unless there is anything to suggest otherwise, (and I haven’t found anything) these systems are not True AI. They are advanced Video Contents Analysis systems (Image Recognition Systems).
Like any technology, when used for the right application and in a favourable environment they can be a step in the direction to enhance pedestrian safety. However, their limitations, as set out above, need to be considered:
1. How comprehensive is the database? My understanding is that most of the technology and hardware comes from a limited number of manufacturers, most of which are in China. I’m assuming they come with a standard database installed but can this be updated?
If the databases can be updated, then there is a possibility that different systems available on the market will have different databases and therefore different levels of effectiveness.
2. If the data base is static, how is it updated, does this require manual intervention?
3. Consider your environment. Pedestrian detection cameras are after all just cameras and rely on a clear field of vision to the object or the pedestrian to detect it. No clear field of vision, no detection. Think building corners, stacks of material, fog, rain, vehicle buckets….
If we go by the definition of True AI, using AI to describe pedestrian – detection cameras is a misnomer. However, AI is a buzz word in all industries, so the popularity of its use in describing all sorts of products isn’t surprising.
I’ll end with a fact from MIT Technology Review: ‘up to 60% of all products marketed using AI do not meet the required definition.’
