Tom Oldroyd: Good morning, and thank you for joining. My name’s Tom Oldroyd, and I’m the Director of Strategy and Sales here at Semantics 21. What we’re intending to do today is show you a product that we released about a year and a half ago, but it’s starting to grow as more and more law enforcement agencies and customers use it and spot gaps in the market that we can very quickly patch.
The tool is becoming one of our fastest-selling products, and hopefully we’ll show you some of its key features and why it’s been designed the way it has. My background is law enforcement. I was a police sergeant at Nottinghamshire Police for 17 years, where I worked in digital forensics with Shona, who’s on the call.
If it had a battery or a plug connected to it, it was my team’s responsibility to look after it. Prior to policing, I was in pharmaceuticals — I probably should have stayed there, there was a lot more money! But policing was a great experience, not only because of the cool stuff that policing does, but because technology at the time was really starting to grow.
I think digital forensics tools haven’t really kept up in certain areas, which is where the team at Semantics 21 has started to plug those gaps. I’ve also worked for BlackBag Technologies over in California, and for Cellebrite, who most of you will probably be using for your phone extractions.
But now I’m working for Semantics 21. S21 is, ultimately, an AI company under the hood. We design various AI products that concentrate on media review — so if it’s pictures, audio or video, we’ve normally got a solution that can assist. The AI capabilities we’ve been creating allow you to review content on your own platform.
It’s completely offline and it’s very cost-efficient. It’s clearly the way forward for us to be able to deal with the huge volumes of data that you’re now receiving for all different types of crime. Although we’re specialists and normally deal with law enforcement, we also have military customers.
And with things like CCTV, we’re starting to see the CCTV review tool used in ordinary commercial industries — theme parks, aero industries, those sorts of places — where they have the same problem as all of us. They need to find the needle in a haystack. So in terms of a Semantics 21 overview, these are the tools that we have available.
If any of you have used products like C4All or C4P in the olden days, those were replaced by Griffeye, which was a bigger media review tool. We’re the competitor to that. S21 VisionX is our primary product. For anything picture- and video-based, that’s the solution to be using, and it’s incredibly quick.
Again, all of the AI works completely offline. We’re seeing most of our customers switch from our competitor over to us, not only because we’re normally more cost-efficient, but because we’re far quicker to process cases and we support more file formats. So we normally give you a better overview. If you’re looking at picture and video review outside of the CCTV world, I’d encourage you to have a play with VisionX — you’re more than welcome to have a trial.
VisionX has a back-end database, so we track exposure to CSAM, which is probably one of the most distressing parts of policing. Even in organisations like HMRC or the Serious Fraud Office, when you’re dealing with digital media, I’m pretty sure you’re accidentally coming across material such as extremist content or child abuse content.
The software is designed so that you can pivot. If you’re dealing with a fraud offence through HMRC and very quickly realise you’ve got CSAM or contact abuse, you can switch the software over, and it caters for both. The back-end database also tracks exposure levels, which is really important for organisations that need to say, “These are my staff, and this is the exposure they’ve had.”
The Investigator Wellbeing monitor tracks that data in the background. Good databases, intelligence and big data are critical to us, so we also have the Global Alliance Database. Again, that’s outside of the CCTV world, but it’s a bit like CAID. I know many of you on the call are from the UK, and things like Project VIC and CAID are very good databases.
They normally hold around 120 to 130 million hashes. We’ve worked with our customers to create the world’s largest media intelligence database, and we’re currently at 3.7 billion hashes. If we took an investigation today — a phone and computer examination — downloaded the media and compared it to our database, you’d normally find that about 80% of that media has been pre-labelled by somebody using our software internationally.
We’ll show you who’s seen it, how they graded it, and whether they’ve said it’s legally held or sexual content. That means you can review content far quicker, which again is one of the reasons why many people are switching from their existing tools to something more modern and state-of-the-art.
Now, you’re all here for the CCTV product, which is what this webinar is about. The reason we created this solution is that we realised VisionX customers were reviewing traditional CCTV footage in our sex offences tool, and it really wasn’t the correct fit.
So we decided to take the engines from VisionX and create the CCTV solution, which is what we’re going to show you today. We’ve also got S21 Transcriber, which was built for the UK’s National Crime Agency. That takes the audio out of video or audio files, and we can transcribe and diarise it.
It saves you having to write out what’s being heard. Let the AI do it — and again, it’s all completely offline. In today’s session, we’re going to concentrate on CCTV, and there are a few areas we’ll cover. The first one is the challenge that you’re all currently facing.
We’ve known for the last 10 or 15 years that digital evidence is only going to grow, and we’re not seeing that slow down. Video evidence is probably one of the biggest challenges we now have, and it’s still growing with technology everywhere — Ring doorbells, dash cams fitted into vehicles by manufacturers. That’s all evidence that’s going to be useful to you at a later date.
We know that’s going to continue to be the case, so we needed to build solutions that could cater for it and help you out. We also know that, especially within the UK, when we deploy digital forensics tools we’ve got to be very mindful that we’re using AI, so we need to comply with the different standards being created internationally.
The new police AI department is looking at what it needs us to do as a manufacturer to make sure you’re confident the tool does what it says it should do — and also at not giving all of that capability to everyone, because in some cases frontline investigators don’t need to use AI.
It may be out of scope of what they can do, so we’ve taken that into consideration. We’re the only company in the world that’s designed the software to have two versions. They look and feel exactly the same, but less technical users have a frontline mode, and more advanced users can have the pro mode.
We’ll also show you some examples of how the software works, because I don’t like just looking at PowerPoint slides — I’d rather show you the product. Then we want you to use the software on your own cases. You may have historic test data that you could run through a CCTV review tool today, and then compare what we produce.
It’s quite interesting: there’s a very popular CCTV solution being used by many UK law enforcement agencies, but one particular force said, “We have some test data, and we’ve never been able to find the murderer.” The murderer appears three times in the videos and they’d never found them with this really big tool. They said, “We can’t use AI because we just don’t have confidence in it.”
We took their video samples and found the murderer every time — all three times. They said, “This is the first time I’ve ever seen AI actually work, and we now have confidence that it works.” So we always say to people, “Use your data. Don’t just listen to our sales pitch. Test it. Really hammer the software.”
We know the explosion in CCTV is everywhere. We don’t really need to teach people in the UK this — we’re probably one of the countries most densely populated with video cameras. But we’re now starting to see law enforcement provide a better level of service by saying to members of the public, “If you have video, you can upload it.”
That’s brilliant. We’ll take that video and look at the poor driving that’s been reported, or somebody using their mobile phone, or antisocial behaviour. The problem now isn’t the technical skill to view the video; it’s the sheer volume. Police forces are having to say to members of the public, “Yes, you can upload the video, but you’ve given us 24 hours of footage because someone scratched the side of your car.
“It’s disproportionate to the investigation. We can’t afford to do it.” Well, technology now allows us to do that, and that’s exactly what this solution has been designed for. We also know that reviewing this huge amount of video evidence is just too expensive, and you can’t keep up with the cost of the sheer volume you’re having to process and review.
Some solutions use online platforms that consume credits and tokens, which are becoming more and more expensive. We know that one of the big AI companies law enforcement have been using has now gone consumption-based, which gets very expensive every time you use their AI.
The idea here is that it’s all offline, so we’re trying to make things work efficiently for you. We want to make sure everything’s verified and that there’s a complete audit trail of what we’re doing in the background. And when we look at the data you’re producing and reviewing, can we provide you with extra niceties, such as redaction?
When members of the public ask for a copy of their own CCTV — say they’ve walked into a bank and want a copy — it’s quite difficult for banks to provide it. Well, not anymore. We can take a sample picture of that member of the public, redact everyone else’s face, and leave a version with just them in it. So for disclosure purposes, the software is really starting to help.
What we’ve created is, hopefully, a single, completely offline solution that uses your hardware, which means you can load these videos in. Now, we’ll be completely transparent: we only support a very small number of bespoke proprietary formats, and they’re normally custom-built for customers who say, “We have a particular video format that no one else can support. Can you guys do it?”
Our engineering teams normally have that capability, but we can’t do it on a huge scale. So what we normally say is use something like DVR Examiner from Magnet — which is now Magnet Witness — to do the conversion, and then we can play the universal formats. Then we can load them and do the searching.
We use AI that’s going to work. We’re not going to promise you a solution that can identify vehicle makes and models, because the solutions out there just aren’t good enough. They’re misleading, and people are trying to push AI a little bit further than it can currently go.
We want you to check and validate. This isn’t about removing the human. This is about filtering — a bit like Ctrl+F on a keyboard when you’re looking for text in an Excel document. We all use that every day. This is about using AI to do these basic searches, and then you validate the results. They’re never going to be 100%, so we have to validate them.
And when we look at redacting, sharing and enhancing capabilities, those are controlled, so that you have trust and confidence in the tool and can push data out for people to use. I mentioned the UK, and I appreciate this video will be shared internationally, but we’re starting to see the same thing happen outside the UK.
We don’t want to provide a one-size-fits-all solution. Frontline investigators with very little technical experience don’t really need the ability to run certain AI models or enhancement capabilities. What we don’t want is for them to say, “Here’s this image of this person. I’ve used AI to increase the quality of the picture, and I’m going to put it on Crimestoppers.”
Or, “I’m going to let members of the public look at this picture that an AI has generated and enhanced.” The facial match could be very good, but certain details may be missing, which could mislead a member of the public who says, “That looks just like somebody I know, but there’s a tattoo missing.”
If the AI hasn’t accounted for the tattoo, it may not show it. So we restrict certain things to people who are trained, competent and understand the technology. That doesn’t stop somebody with the pro mode, who has the expertise and does the AI processing, from providing a package back to someone on the frontline to look at the results.
The frontline users just don’t have the ability to run the AI up front. So for things like ISO 17025, or the Forensic Science Regulator’s Code of Practice, frontline investigators have a very basic review tool for capturing frames, a quicker way of reviewing using motion analysis, and the ability to create reports. That removes a huge demand from the technical people and leaves them with the pro tool, which has the full AI and enhancement capability.
If you need extra staff — say we have more riots, which is one of the reasons this software has been so successful — you can scale up. When we had the riots, forces needed to very quickly upskill people, and major crime teams said, “We have five or six people using your software, but we now need 15, because we can’t deal with the demand of 15,000 videos to review.”
So we can take users on the frontline version, give them access to the pro version and apply the additional training, but the software looks and feels the same, so there isn’t a big learning curve. When we look at our internal development team, we want to make sure everything we develop is cop-proof.
If a police officer can use it, that’s a pretty big tick in the box that we’ve designed something really useful and easy to pick up. So let’s look at some examples. I’ll show you how simple it is to run some really basic searches, and please feel free to ask any questions in the chat.
Charlotte will keep an eye on the chat messages, or just unmute yourself, jump in and ask me, and we can steer the session to fit your requirements. One of the basic things we see is huge amounts of video being brought in, and we’re normally intelligence-led — we know what we’re looking for, we just don’t know where it is.
In one case, one of the Welsh police forces that uses our software had 600 hours of video for a homicide that they needed to review. The offline processing time was, I think, about six or seven hours for all 600 hours of footage. Then, very quickly — in less than an hour — they were able to find the person of interest. That’s the sort of speed and capability the solution provides.
So let’s jump into the software. I’ll go back and forth between the software and the slides, but I’ll show you what the product looks like. If you’ve ever used any of our other products, S21 CCTV looks very similar in layout and functionality.
We do this for the same reason Microsoft did. We’re all very familiar with Microsoft Word, so when Microsoft gave us Excel and PowerPoint, they became second nature because of the shared design. We’re doing the same thing: if you’ve used one of our other tools, switching to S21 CCTV or another of our solutions should feel the same.
We have pop-out bars along the left or right. Whether you’re left- or right-handed normally changes where you want them, and you can have different bars on different sides. The functionality is there because we want to make sure the software fits around your needs.
Across the top, we have some very basic options: creating cases, adding evidence to an investigation, reprocessing, and all our normal settings and save options. We can go back to our grid, which shows all the video files that have been added.
We have a multi-sync player, which we’ll talk about shortly, that allows us to play multiple videos. Our AI options and some of our search options are across the top here. If you were in frontline mode, you wouldn’t see these, because you wouldn’t be able to trigger the AI.
But if you were provided with an AI package that had already been processed, you would see them. Over here, we have a notepad where users can keep their own contemporaneous notes. They’re not notes that would normally be reported on — it’s just a scratch pad for jotting down what you think the suspect may look like or writing quick descriptions.
We have our reporting options. We can export to S21 Transcriber, export videos or segments of videos, and see our process summary, which shows everything the software has been doing. In the top right-hand corner, we have our quick-start video guides and our written guides.
We’re also the only company within forensics that I’m aware of that provides WhatsApp support. If you ever need to, you can scan the QR code and talk straight to the S21 team about anything to do with the application. It could be very basic — “How do I create a report?” or “Can I get a new licence key?”
The team will reply instantly or phone you back directly on your WhatsApp number. We want to reduce any friction when you’re dealing with us as a company, and make sure the product is fit for purpose. When we load our videos in, you’ll notice they’re all on the grid, and we can very quickly scan over them to see the context of each video file.
The first scenario is: how do we find a particular person in a video? There are various ways to filter these videos down. If I want to, I can filter to just the videos I’m interested in, or I can run global searches across all my video sets.
It could be that I know this is the video of interest here across the bottom — just some street CCTV footage — and I want to try to find someone relevant in it. There are various playback options available, all controlled from these top ribbon bars.
You can see the video playing in the background, but I can increase the playback speed or slow it down. I can flip the video if it’s been recorded in a mirror or otherwise flipped. We can change the colour levels, hue, saturation, contrast — all the very basic things we need.
We can also split the video into multiple segments. This is the same video — it’s about 25 minutes long — but you can see we’ve now split it into eight equally spaced segments. If I need human eyes to review the entire video, I now can, and I can speed it up as well.
That means we’re not relying on any AI at all; we’re just making things a little more efficient in how we display the video. Now, the task is to find a person of interest. As I say, we’re normally intelligence-led. If a member of the public says, “My car’s been scratched,” we could draw over the car.
If we’re looking for people of interest and we know they’ve walked down this particular footpath, we can draw over it, and on the top toolbar we can run what’s called motion analysis. It’s targeted, so we’re focusing on this specific zone of the video. The 25 minutes of video should only take a few seconds, because this is motion analysis.
It’s not using AI, so even a frontline investigator would have this capability. You’ll notice little orange markers start to appear along the bottom toolbar. Those orange markers mean that somebody has been detected walking into that specific zone. It doesn’t matter whether the video is 10 minutes, 20 minutes or five hours long.
It really makes no difference. If you’ve ever dealt with certain CCTV systems, you’ll know that video spanning 24 hours is created in smaller chunks. The reason is that if the system crashes 15 hours into the recording, it doesn’t damage the entire video — you’ve got those smaller chunks as backups.
You can bring all of those videos in. Maybe it’s a 24-hour video split into 48 clips. You can bring them all in and treat them as the same scene. When you draw over the footpath once, you can tell the software to treat all of the other 48 videos the same way, so you don’t have to keep drawing over the front door to see who’s come out of the house.
That’s a really useful feature. The team also thought that, instead of users having to drag the bar along with the mouse, we could use Smart Play. Smart Play plays the video where motion happens and then jumps to the next segment. For many investigators, the idea is to find the relevant parts and then flag them very quickly.
If we think this is a potential victim, we can just hit the Victim button at the bottom. It captures the frame and tags it as the victim. If the next person to come in is a potential witness, we hit the Witness flag, and again it captures that frame under the witness tag.
When we come to create our reports, we can say, “These are the people we believe are our suspects,” or “These are our victims.” It’s a really simple workflow, and again, it’s completely audited, so we know what the system has been doing. That’s how we can detect people walking through a certain area of interest.
If we’re looking at audio, we can very simply run our audio analysis, and a blue bar appears along the bottom of the video. This is really useful where you have covert recording devices, a voyeurism camera, or an office camera where nothing’s happening on screen but something may be happening in the background.
It allows us to say there’s audio in the background — here’s a spike with a gunshot, maybe breaking glass, or someone talking behind the camera. That can be really useful when you’re not seeing anything visually. Now, because we’re using the pro mode, there are some capabilities shown here that you wouldn’t see as a frontline user.
One of them is stabilisation. Lots of investigators use body-worn cameras, and we’re even seeing them in supermarkets now. They’re really useful for capturing data, and they’re definitely doing their job of reducing complaints and the amount of conflict law enforcement have to deal with.
But when you’re moving around with a camera, you get a lot of jitter and movement, and you can lose detail. We looked at what companies like GoPro have done with things like HyperSmooth, where the device compensates for movement while recording.
Well, we can now compensate for that movement after the fact. We can take body-worn camera or dash cam footage and stabilise it to remove the bounce and movement. It can uncover some really useful clues. We found a really good example online of a police shooting in America where the officers were wearing body-worn cameras.
They shoot the suspect, and he dies from his gunshot wounds, but you can’t really see what the suspect was doing. When you stabilise the video, you find more detail showing that he has a firearm and is actually opening fire on the officers, and they’re returning fire.
So that’s just a different way of using technology to hopefully uncover additional clues. We also have our facial redaction system, and here we have an activity solution — I’ll show you a better example of how we can target not only movement in a particular area but also how we use our AI.
We may not need to use AI on an entire video; we might want to use it on a specific part, which is really quite clever. I’ll show you how we do that in a second. So far, we’ve just run a very basic motion analysis to find people with no AI at all. But what happens if the scale of your investigation is far bigger?
What about the riots two years ago — and we now seem to have them every few months? We can capture that data, but it’s about the review and how we find the critical element we’re looking for. When we looked at the riots a few years ago, law enforcement were telling us the review time for those cases was two and a half to three years.
We provided the solution to law enforcement for free through the NCA and said, “Look, we want to help out. We know UK police are struggling. Please use the tool for free, let us know how you get on, and give us feedback so we can benefit from it.” The forces that took us up on the offer took a two-and-a-half-to-three-year review down to three months, and they were blown away by the efficiency savings.
Strangely enough, after the incidents, they’ve now bought the solution and deployed it. The forces that didn’t take us up on it did take two and a half to three years to review, and they’re now realising they missed a trick. They’re now buying the solution, and it’s saving them time on cases going forward. So it really is working well.
Now, how does the AI work? If we’re looking for a specific object or person, I just click the simple button at the top that looks like a little person. This processes all of my videos on the left, and you’ll see some are orange. Those are flagged because the system is saying people appear in those video files, and as I scroll down, you’ll notice the orange markers change.
Instead of drawing over a specific area, it may be that I just need to look at people. I know someone’s committing an offence, I have the offender’s description, and I want to go through the video files to see if I can find a match. This will show you any sign of a person or face.
Whether the person is wearing a hijab, a balaclava or a crash helmet makes no difference — if we can detect the body parts, the system will show you the results. We know video analysis can take a long time, so you can even turn on auto-scroll and go through your video files while the screen scrolls automatically.
I know on Teams this can look a bit stuttery, but on my laptop — and this is only a £1,000 laptop, nothing high-end — it’s completely smooth. I get to see all the results as they come through. I can simply click the stop button, and if I need to look at someone’s details, I can click on them and see the frame where they were detected.
This shows an officer here, and again, it’s quite poor quality. It isn’t good enough to do facial recognition, but it is good enough to say that’s a face. It may be that the person’s clothing is enough — you’re looking for someone in a green top, probably something like a North Face jacket with a grey hood.
That could be enough for you to say, “We believe that’s our suspect.” Even though we’re looking at the face, we can’t recognise it because the detail is too low. So we’ll show you both body parts and faces. If you were looking for a specific person’s face, you could say, “We don’t want to see body parts. We only want to see faces that have been detected.”
We tell the AI to start filtering based on what it can see and understand, and this is all based on AI confidence. Because we build our own AI, we can provide capabilities that some solutions can’t, because a lot of solutions have to source their AI models externally.
They’re using things like Hugging Face or GitHub, where people have done the research, created a model and put it online, and then someone builds it into a forensic solution. Because we create our own, we can provide things like AI confidence. I’m just going to zoom in slightly so people on the call can see it. All of our products do the same.
The AI confidence down here reflects the fact that when the AI looks at the biometrics of a face, all it’s doing is creating a calculation. We then look at whether there are similar calculations that would indicate a potential match. It’s very simple — it’s maths. That’s all artificial intelligence is really doing.
Here we’re showing an AI confidence of 80%. If I lower the confidence level to 25%, you’ll notice some of the faces it detects are really poor quality, but at 25% it believes that’s the face of somebody in this office, right over here in the background.
So it has done the job and it does work, but because we’ve lowered the confidence to 25%, we might start getting false positives creeping in. The idea is that if you want more accurate results, you set the confidence level a lot higher, and then you can lower it gradually.
That’s where you’d use the notepad to say, “I’ve done the search at 80% and can’t find the suspect. I’m narrowing the time parameters to a certain window and lowering the confidence level to see if I can find a potential match.” Finding a face in a crowd is going to be incredibly difficult, because what we’re showing you here is every face the system has detected.
The AI can also start to group people, again using that confidence level. On the left-hand side, in this little Faces tab, these are people the computer is starting to group because it believes they’re potentially the same individual. Whether they’re in the same video or any of the videos in our demonstration, it will group them across multiple files.
Once you’ve grouped someone and can see this is potentially the suspect of interest, you mark them as identified and give them a name. If I just make a name up — I have no idea who this person is — and say this is Dave, that creates Dave.
What it’s actually done there is grouped him with somebody already in the system called Dave. If I made that mistake, like I’ve just done, I can select those faces and remove them from the group to say, “No, that’s not actually Dave.”
I can then give him a different name — say, Tom — and it will group him under Tom. All you’re doing is working through the groups of faces the AI has detected. We can run face and body part detection over all our video files and start grouping people using the AI’s own grouping capability, where it believes it’s found a match at a confidence level of, in this case, 83%.
We can verify it by simply opening it up and looking at the frame that’s been detected. We can see this guy has a black top on with a white collar. We can see him in the video again three minutes later — black top, white collar, the same glasses. It’s probably the same person, and the facial recognition has worked within this crowded scene.
Again, we give that person a name. The idea of using names is that we can then filter — “Show me all the pictures where we’ve found Tom,” or Dave, or anyone else. They go into this section called Identified, so Identified Faces are the ones we’ve given names to.
If you think of how Apple Photos does something very similar with pets and family members, it’s exactly the same idea. Once we’ve identified somebody, it gives us extra capabilities. With redaction, for example, we can say, “We’ve found this face, and we want to redact everyone else’s.”
We want to blur everyone else in our investigation but keep, in this case, Dave and Tom in focus. Their faces won’t be blurred. Gone are the days of having specialists use Adobe Premiere to blur out videos — this solution can now do it for you, and we’ll show you an example on some other material.
You might also have a photograph — an external image — and you can add any external image you want to your analysis. When you add it, the system will analyse all your video files to find potential matches, and those potential matches will be displayed on screen.
Again, validate and verify that they’re correct — that the circumstances in which the person is found and the clothing they’re wearing match. Trying to find a person in a crowded video can be incredibly difficult, especially at a riot scene or a football match. There may just be too many faces.
It’s like Where’s Wally on steroids, because you’ve got constant movement. This will outperform human reviewers. We’re not saying human reviewers are poor; we’re saying human reviewers suffer from fatigue. We’ve run various tests with different student groups.
We’ve done it with our internal staff, and with colleagues I used to work with in policing, where I’ve said, “Just give this a go. Watch this five-minute video sample and try to spot this person.” It’s very difficult for them to do.
But when you apply the AI, it can do it at scale. It’s very good at finding faces in groups, and hopefully it prevents that missed opportunity — the missing person or the suspect you’re looking for in a crowded scene, we can detect them and you can filter them down.
We’re not saying remove the human; we’re just saying this is a capability we have, and it’s now been tried and tested. Talk to us offline and we’ll tell you who they are, but there were some prosecutors who were very nervous about the use of AI facial recognition. They then realised that the AI was far better than human reviewers.
Their concern became relying on human reviewers alone without using computer technology, because they realised humans were missing things — and you don’t know what you’ve missed until someone points it out to you. So those are the different ways of using the technology to find people.
We can draw over a certain area, we can detect body parts, and we can look for faces, and the results are shown really simply on the grid. The upcoming version, coming very shortly, can also look for associates. If we have a person of interest, we can look at the footage and say, “There’s the person you’re looking for, but you may also want to look at these people, because we believe they’re associated with them — they’ve been seen together multiple times.”
That gives you additional forensic clues you might not pick up as a human, because the computer can tie things together for you — and we do that with vehicles too, which I’ll explain shortly. So for the scenario on our slide of trying to find people in a case, it’s really easy: we can draw over a certain area to look for motion, and we can run our facial analysis to find people.
Now for our next scenario. Because we have an explosion of cameras across most areas, we normally have multiple cameras looking at the same incident from different perspectives. If you look at body-worn cameras that some of your staff will be wearing, you’ll get a completely different story from different angles.
We did some work with one of the terrorism groups over in the States. When we were talking to them about reviewing cases, they said one of the hardest parts of a bombing at a very big event was seeing the perspective from different cameras at the same time. It’s really useful to say, “We know the bomb goes off here, but we want to see the other cameras at the same moment so we can start tracking the suspect who left the package.”
We can do this using what we call the multi-sync view. The idea is that videos may have different start times, and you want to re-sync them so they play in line. At the moment, we can choose nine screens, four, two or one — it really makes no difference to us.
These could be four body-worn camera videos. Or let’s say we have a stairwell, a passage, an office and a reception area — we can add those cameras in and see four different views of the same property.
In this demonstration, I don’t know how these videos are synced, and I’m pretty sure it won’t be spot on. But I can set the time and date. The system’s time and date may be incorrect, but I can say this one is 18:06, set this one to be pretty much the same, and this one the same. Then let’s make this one slightly different — five seconds different.
You can see this one will now run five seconds later than the other video files. Once all the video files are in sync — it could be an officer-involved shooting, a warrant where you want to see who went through the doors and when, or, in this case, an office — we can play all the videos in sync.
The first three will run, and when we get to the purple bar, the next one kicks in. So we can line all our videos up in order. In a real scenario, we’d see the different times people walk down corridors, the route a person has taken, and what everyone else was doing at that particular time.
If we were interested in someone walking down a particular corridor, we could look at the rest of the people in the office to see what they were doing, and then approach them: “You were sat at your computer when someone walked round that corridor. Did you see what they were carrying? Did they have the stolen laptop with the confidential information? What were you doing at that time? You looked like you were on a voice call — what call was that?” So using multiple camera angles like this is quite clever.
We did this recently with an officer-involved shooting, where an officer said a particular thing had happened. We had the other body-worn cameras and the dash cam from one of the police vehicles, loaded them all into the scenario, and synced them so the times were correct. Then we could see the police cars arriving.
We could see the officers getting out with their body-worn cameras, and we got perspectives of the shooting from different angles. That was really useful, because although one of the officers said he didn’t see something, he did capture it on his body-worn camera, which backed up what the other officer was saying.
When you export that as an overall video to give to the prosecutor or the IPCC, or whoever it may be, you can say, “This is what the scene actually looked like. We’re not relying on just one video file.” So it’s a different way of using video — not so much for searching, but to give you a good overview of how the data fits together.
That’s just another simple feature that one of my customers asked for, which we then built in. Now, one of the things we know you’ll all be looking for in investigations is vehicles. We wanted to give you a very simple way to search for vehicle details — the colour of the vehicle, the type of vehicle, or, luckily for us here in the UK, the registration plates on the front and back of our vehicles.
It’s very simple. Within the application, in our grid view, next to People we have Vehicle Search. Simply clicking the Vehicle Search button turns on all the options to filter your grid. Across all your different cameras — whether it’s a covert camera, camera car footage, an ANPR camera at the side of the road or a sentry camera — you can see all the vehicles detected after processing, by the types we’re looking at.
Again, it’s all built on AI confidence. If I drag the AI confidence down to, in this case, 41%, you’ll see the grid refresh and show me more vehicles that have been detected. If we look at this case here, it’s detected a vehicle in the far distance, and it’s 41% confident that it’s detected that particular car.
We can go really small, and this is where you have to be careful, because you may get some false positives creeping in. But if we need to verify a particular vehicle, instead of looking at the frame, we click it, it moves the video to that point, and the video plays, showing the vehicle as it drives away from us on that camera.
The same options apply. Across the top here we have the dropdown for vehicles: cars, buses and trucks. If I’m interested in bigger vehicles such as trucks, you’ll notice underneath we have our colour search area and our ANPR results.
Then we can say we want to look for any bigger vehicle where the AI believes, at 41% confidence, that it’s a truck. This is where we can mix and match our searches. You’ll notice the ANPR results filter down too — it will only show ANPR results for vehicles categorised as trucks.
Or I could say, “No, I’m looking for all vehicle types, but in a specific colour,” and run the search. Across all my video files, it will now show me any vehicle where we’ve detected the colour red.
So we can filter by vehicle type and colour, and we can even choose which videos to show and search. Then down here we have our ANPR. If I clear the search and I’m looking for a specific VRM I’m interested in — say I search for my own car — you can see there are two hits.
I can confirm the frame that’s been found, and if I double-click, it shows the video at the point where it detected the VRM. We can see, from our made-up camera car footage, the vehicle being parked. So again, it’s just a different way of using the information we have.
Some people have asked us, “Why wouldn’t you provide this capability in frontline mode? Why have you restricted AI to the pro mode?” The reason we lock out some of the AI is that something like colour searching only works on videos that have colour.
When we were watching how investigators used the tool, we’d say, “Play with it — we want to see what you’re doing.” In a few cases, people said, “I’m looking for someone in a red jacket,” or “I’m looking for a blue car, and I’ve run 24 hours of video and there are no hits.”
Well, at six o’clock, because it was getting dark, the camera switched to infrared mode, so it’s black and white and colour searching isn’t going to work. It’s these basic things that we should be thinking about and picking up on. Some of the non-technical users relied on the technology too much, when it was never going to work on black-and-white footage.
That’s why certain features are restricted to technical users, who really should spot that the video is black and white and colour searching won’t work. So that’s how we search for vehicles. I mentioned previously that we’re doing associate detection for people.
We’re also going to do convoy analysis for vehicles. When I was at Nottinghamshire Police, we used to have quite a lot of cash-in-transit robberies, where cash delivery vehicles would be attacked and the money stolen. What we knew was that they were normally put under surveillance for quite a few weeks beforehand so the offenders could work out routes and patterns.
Or look at harassment cases: the person is normally followed not just once but multiple times. If we load in video files of a particular street, we can say, “This is the VRM of interest,” and then look at the vehicles constantly seen trailing that car, which would suggest it’s being followed.
So associate detection for people looks at people walking and being seen together, and convoy analysis for vehicles looks at which vehicles are following other vehicles in a way that might be suspicious and warrant a closer look. Now, when you’re looking for specific vehicles in video footage, you could be overloaded with too many hits, and we’ve taken that into consideration.
If I look at a scene like this one — a motorway with three lanes — it could be that we know our suspect vehicle is travelling in lane three. You can run a targeted motion area and then, in this little search dropdown, tell the software whether you’re looking for faces, people, vehicles, scene activity or ANPR.
If I say I’m looking for all vehicles within that area and run the search, it will now only show me vehicles detected in that specific zone. If you’re working at an airport and you’re interested in people walking through a particular corridor, you could draw over that area and say, “Just do facial matching for people here,” to cut down the search parameters.
So the AI can be used in various ways to help you screen and filter as quickly as possible. And again, we’re keeping to the simple things we know the AI can detect reliably. We’re not looking at makes and models — we’re not going that bespoke — because when we’ve tested the modelling systems available, the results are too poor, and we don’t want to give you something that just isn’t going to work.
So that’s how we find and narrow down vehicles using colour, vehicle type and size, and ANPR. Now, when you’re preparing footage for a public appeal or for disclosure — for example, members of the public who enter your site have the right to request any footage you’ve captured of them.
It happens a lot in commercial industries now, where people leave their employer and then make a Freedom of Information request or a subject access request (SAR) for all the footage they’ve been captured in. You can use our facial recognition technology to say, “This is the person who no longer works here, and they want a copy of all the video they appear in.”
You can detect them very quickly, then use the facial redaction system to create that version. I’ll show you an example as a pre-recorded video, because the process runs at a one-to-one ratio. Say we’re interested in this particular guy carrying the sign.
We’d select the person of interest and say, “His face needs to stay in focus, and everyone else in the video needs to be blurred automatically.” As I say, it’s a one-to-one process: if the video is an hour long, the redaction will take an hour. The reason it takes that long is that we constantly run checks on the video file.
We’ve done some work for another company reviewing the capabilities of other redaction tools. We found that if we went frame by frame, the redaction would drop off and we could actually identify the people behind it. We’ve created a solution where that shouldn’t happen.
It’s constantly checking the frames, and we make sure the blur is applied across every frame. So you can be confident that if you’ve blurred a video with our software, it will stay blurred. Nobody will be able to go frame by frame to find where the blurring has fallen apart.
That’s why it’s a longer process — it’s one of the downsides, but it just runs in the background. You select the face, the process runs, and it creates a file for you. Then, again, verify and validate all your results and findings. So those are the key points we wanted to show you about the software.
It really is simple to use, and so far we’re getting that big tick to say it’s cop-proof. The software includes built-in videos, so if you want to learn any of the features, you don’t have to go on a paid training course. You can just watch those videos in the application, and they’ll teach you all the basic things we’ve discussed today.
What we’d like people to do is not just take our word for it — and we know you won’t. We’d like to say to people, “Take a trial of the software and install it on your own infrastructure.” Test it out and give us your feedback on what you like and what you don’t. If there are bits we’ve got wrong, please let us know.
This is still quite a young piece of software that we’re developing, and we’re constantly getting feedback from people who say, “We love it, we’d just like it to do a couple of other things.” We can build those in for you and provide that additional support. Do a fair test.
Like I said, not too long ago we worked with a police force that had a homicide sample, where they as humans had found the suspect. They said, “We were really surprised that the biggest company, the one everyone seems to be switching to, missed it. They missed the murderer three times.”
They said, “If you can’t pass that test, you don’t get your foot in the door.” Luckily, we passed the test — we found the murderer all three times — so they were quite impressed. We’ll always look to make things better and easier, but we want you to test and really thrash the software and not hold back.
Use the same footage you use every day. Maybe you’ve worked on historic cases where you’ve already provided results — test the software again and see if you get a different result, whether it’s quicker, and whether there are savings we can find for you. And again, we stress: validate your findings.
Don’t rely on just what a computer is saying. This is filtering technology — that’s all our AI is designed to do. We can filter down a huge amount of data, and you have the transparency of filtering by confidence level. We’re not going to lock it at 90% confidence and hide everything else.
We think that’s completely wrong. We don’t agree with that way of working, and we’d rather give you the ability to change it. We also find that if you already have CCTV video tools, there’s normally a cost saving in switching to our product. We’re a smaller British business.
We’re not one of the big US-style firms with bigger overheads or export fees. We’re based here, so we can normally be very competitive on pricing. If you’d like a copy of the software, please reach out to me or any of the team at Blue Lights Digital, and we can help you out.
There’s a QR code on screen if you want to scan it. It’ll take you to a contact form where you can request a trial copy of the software. The trials aren’t limited to seven days — we can extend them so you can really try it out and make sure you’re confident the tool does exactly what you need before you go ahead and purchase.
We also have some really friendly UK law enforcement customers who have agreed to speak to people. A lot of people don’t want a sales pitch; they want an honest review. We can put you in contact with customers who use the software day to day, so you can hear their findings and how they’ve got on working with Semantics 21.
If there are any questions, please feel free to shout up. We can also walk you through any parts of the app that I may have missed or haven’t covered in enough detail. It’s up to you, so if anyone has any questions, please fire them across. Did anything come in while I was talking, Charlotte?
Charlotte Husselbee: Nothing, no.
Tom Oldroyd: That either means people aren’t there or I’ve covered everything!
Charlotte Husselbee: Oh, we have got a question. Liana, over to you.
Liana Kohli: Hi there. It looks like a really great product, so thank you for showing us. My question is — because I come from a quality background — can you explain how the frontline and pro modes align with the NPCC activity levels one, two and three in their Framework for Video Evidence document, please?
Tom Oldroyd: Yes. With the frontline mode, we’re compliant with the NPCC guidance on providing a player. Ultimately, what we’re saying is that this is a player. We know people will naturally use things like VLC or Microsoft’s built-in media player — something that doesn’t provide any additional enhancement capability, or anything that’s out of scope, like AI.
Any enhancement would take it into the higher levels. Frontline mode is simply a player — that’s all it is, with a wrapper that makes it look a bit more police-oriented. That’s why, when we sell it, it doesn’t really cost any extra. When we sell the pro mode to enterprise users, we say, “The frontline mode comes as a by-product. You’re really buying the pro mode because it has the geeky stuff.”
The frontline mode is a nice-to-have that provides that basic review. It takes into account the NPCC guidelines: it must be simple, and it mustn’t go out of scope to where people can start interpreting evidence or making changes.
There’s nothing in it that would count as an enhancement capability. It literally lets users capture frames, and they can generate a report if their agency allows it — it may even need to be signed off. All the auditing happens in the background as well. What we’re currently working on with South Wales Police is what level of auditing they want for the AI side.
We can go very granular. When we spoke with the NPCC and the Forensic Science Regulator, we always asked, “Would you like us to become accredited? Is that something we can do as a vendor?” Within our other solutions, we build ISO validation checking into the software.
They said, “You can’t become accredited, because 17025 is about how the software is deployed and used. But if you can create capabilities that make that testing easier, we’d really appreciate it.” Law enforcement doesn’t want to keep regenerating testing. We understand our software better than anyone else, and if we find faults, we’ll fix them.
Software will have issues, but if you have an easy way of testing it, that’s exactly what we’re designed to support. So we appreciate that the player is basic — it replaces VLC and stops people using Microsoft products and VLC, which aren’t really fit for purpose. It’s a purpose-built transit tool at that low level, and then the AI goes to the higher end.
That’s where we’d say it goes to an accredited lab, where the specialist teams understand how the technology is being used and what the AI is actually doing under the hood — for example, not using colour detection for vehicles on black-and-white footage, because you won’t get any results. There are also certain features we haven’t shown you today, like clothing detection.
We’ve even hidden those away in the application, because even in pro mode it’s a feature we expect someone to deliberately turn on, as it can make mistakes. If you’re looking for blue jeans and the lighting’s poor or there’s sun glare, the results are going to be wrong.
So yes, we definitely take the different levels into account. And again, we can put you in contact with law enforcement customers who have already been through those pains and can tell you, “This is what frontline gets, this is what major crime gets, this is what the TSU gets.”
Liana Kohli: Brilliant. Thank you for that.
Charlotte Husselbee: Thank you. Anybody else have anything else? No? Just quickly, because a few people joined the call after we’d started, I’m about to put my email address in the chat. If you haven’t had an email from me with the link for today, just drop me an email so I’ve got yours and can send you the recording. If it would benefit somebody else in your team, I can send it over to them as well.
Also, there may be other questions once we hang up today, so save my email address. You can contact me any time, even just for a chat — I love a chat! So anything at all, just let me know. I think that’s it. We’ve done it — one minute to go, and we’ve done an hour. Well done.
Tom Oldroyd: It’s like we’ve done it before!
Charlotte Husselbee: It’s like you’ve done it before! Okay, we’ll stop recording now, if that’s all right with everybody. Lovely to meet you all, and do reach out if you have any other questions.
