 Cool, Chiara Koto, something a little bit different, not medicine at all, but engineering and social sciences. So I'd like to share some of my PhD research with you tonight around creating a practical and ethical smart video analytics system. And you might be thinking to yourself, what is a video analytics system at this point? So a video analytics system is essentially a system that analyzes video. So it's really all in the name. But the idea is that we want to try and get some high level information about what's happening in that video. So let's take a supermarket, for example. We might be able to use camera footage that's taken from inside that supermarket and find out where the people are. So we can detect the people. We can count the number of people inside the supermarket. That can help inform some insights into how people are using that supermarket, so which pass they're taking, how many people there are at certain times of the day that can feed into high level business decisions around which items do we need to put closer to the entrances and exits? How many staff do we need at certain times? And we can get some better insights into consumer behavior. So you can start to answer questions like which items are people picking up but not necessarily buying so that we can send out a special offer next month. And so you might call video analytics a form of video surveillance. But this motivating scenario shows that we're moving beyond sort of law enforcement and public safety type applications. So as the costs of computer vision start to go down and the accuracy of these systems keeps improving, we're going to see more commercial entities like supermarkets try to use this type of technology to give them some more business intelligence. And so the primary enabling technology behind all of this is artificial intelligence and machine learning. So there have been a lot of convolutional neural networks that have been developed over the last decade or so that have really accelerated the development in this area and they've achieved really high levels of accuracy. The catch is that most of these algorithms while very accurate are very, very slow. So there's a recent algorithm that is designed for person detection and it operates at about 90% accuracy in unconstrained environments. So it's used in airports and train stations and it can get 90% accuracy but it takes six minutes to process one second of footage. So you get this massive backlog of footage that's not being processed. And so we need to find some solution to that. In addition to that, we have embedded systems that are being used in the real world. So these are computers that are resource constrained. They're not as powerful as the computers that these algorithms are being developed on. So we don't have $1,000 GPUs to use. We have very small computers that we need to do this processing on. And so as is common in engineering, we have to find a balance between speed and accuracy. We need enough speed to be able to do this in real time but also if the systems are inaccurate then there's sort of no point. Another important thing to remember is that in the real world, we don't often have labeled training data. So what that means is that for a lot of these algorithms that are currently being developed, you need hundreds of thousands of examples of the types of images that you're trying to process. So if you're trying to do facial recognition, you might need millions of examples of people's faces to know what sort of facial features to look for. But if you're in a supermarket and you've got a customer who's coming for the first time, where are you going to get that training data? So we need to use what we call unsupervised learning algorithms instead. And so the bulk of the work in my PhD is the development of this multi-target, multi-camera tracking system. So this system was developed at a test environment over at Newmarket Campus where we've got a couple of cameras that are overlapping and watching the same space. And so we've used this plug and play modular pipeline which allows us to choose a couple of different algorithms and use them at different points in the system to try and balance between speed and accuracy. And while we used a couple of off-the-shelf algorithms, it turns out that not all of the algorithms out there are any good. So we had to develop a couple of ourselves. And I'm using the word we here. So actually I had to develop a couple myself. And so we've got three novel algorithms at the background estimation, feature matching, and position tracking stages designed to try and balance speed and accuracy. And this system uses unsupervised learning. So we don't need a bunch of training data. These algorithms are all selected to meet that criteria. And in particular, this system's going to keep improving over time. So as it operates, it's going to keep learning about how to process that footage and get better. And we've done experiments that show that you can run this on embedded hardware in real time. So here's a video. So here you can see the system operating. On the top right you can see a map. So these dots are being drawn as I move around this space. This top wall here is this wall on these two cameras. And you can see that as I'm walking around, it's drawing the dots in real time. And this system's running at about 20 to 30 frames per second. So that's pretty fast on a standard PC. And even though it's not 100% accurate, you can see that it's picking me up enough of the time that this path that's being drawn here is pretty representative of the path that I'm taking as I move throughout the room. But while developing the technical system as the vast majority of my PhD, I sort of feel bad about it, right? Because as probably most of you are thinking, it seems a little bit Orwellian, seems a little bit big brother. And it seems to me that a lot of engineers sort of just build the cool thing and then worry about the consequences later. So I wanted to really understand the ethics behind my work and to place my work within that context. And so I've done a bunch of work into privacy. And we wanted to understand what makes people feel more or less comfortable about the use of surveillance cameras. So we ran a massive survey, well, reasonably sized survey depending on which field you're from, from our field this is pretty big. And it's a mixed methods online survey where we asked a bunch of scenario-based questions. So what we did was that we told people stories, right? We said, here's a story, an example of how cameras are being used. All of them were sort of practical and potentially real scenarios. And then asked participants to put themselves in that context. And what we found was sort of five major factors that pushed people one way or another. So access, who has access to that video footage? Is it one government official or is it a thousand corporate employees? Is there a person in the loop or is all of this footage going to be processed automatically? Can the person who's being observed actually be identified or more importantly will there be personally targeted actions as a result of being observed? Does the purpose of having that surveillance camera system benefit the person that's being observed? And last of all, do we trust the owner of the network? Do we believe that they're competent? So a lot of people said that they didn't believe that the government or corporations could be trusted to hold on to their data securely. So what do we do with that? Well, there are two pathways to actually implement some of the relevant protections. So we can do this by regulation. Governments must pass laws that protect our privacy and enforce them by punishing those that infringe upon the privacy rights of individuals. Sounds great. It's probably what we expect. The problem being that governments are really, really slow and they tend to be well behind where the technology is. So our other option is to do it by design. So technology developers like myself should or must build privacy into their systems so that it is impossible to infringe upon people's privacy by default. So here's one that I designed earlier. So this is the privacy affirming architecture. And what we're doing here is that we're doing some of the processing at the point of image capture. So we have smart cameras that are able to do some of that processing and we're separating that from the central coordinator or the central server. And what happens here is that we can do some preliminary processing and then delete the footage. So the end effect is that the human user who is using it from the central coordinator never sees that footage. And while this doesn't solve all of the privacy problems, it does go some way towards protecting privacy by default because what that means is that if you've got malicious system owner who is a little bit pervy and is sort of voyeuristically looking at the customers as they move through the supermarket, if they have the system, they can't because the footage doesn't exist. And one thing that you'll notice is that it supports that modular pipeline that I described earlier. So these two elements actually really go hand in hand. And the other thing that you'll notice is that this is a technical solution. So this is a solution that no lawmaker could come up with sitting in parliament by themselves. This is something that you have to do with technologists in conjunction with policy makers. So to sort of conclude, I've been talking about my thesis and the system that I've created and it's two things. It's practical in the sense that it can run on embedded systems in real time and it's ethical in the sense that it actually does care a little bit about privacy and does something to try and protect people's privacy in that sense. And those are the two factors that we really need to take into account when we're trying to translate a lot of this academic research into the real world. Now, mihi nui kia koutou katoa. I'd like to thank my colleagues in the embedded systems research group, my supervisors, Mortessor and Kevin, my good friend back to the robot who plays a good game chess. And all of you here for listening tonight. I'm happy to take some questions.