Quang Trinh of Axis Commutations presents how AI is going to transform the security camera market.
Quang Trinh:
How’s everyone doing today? My name is Quang Trinh and I am the Manager of the Professional Services team here at Axis Communications. For some of you that don’t know who Axis Communications is, we’ve been around for a while. We started in the ’80s doing print servers for the IT industry. And then in 1996 we actually invented the first IP camera, an actual device on the network running by itself without needing to be connected to a computer with a camera sensor. And then from there, we entered the security industry market where a lot of things were analog.
Quang Trinh:
And from that time to now, we’ve been really trying to push the industry into moving towards accepting IP video and now it’s a norm. A lot of different organizations out there see IP video as the future of their infrastructure. And the most interesting aspect of us when we invented the IP camera back in 1996 was at that time we realized that there was not a lot of processing power that you can put on the edge to actually encode the images or the videos properly. So what that led to was us developing our own chipsets.
Quang Trinh:
So basically we developed an ASIC processor design to just specifically do video encoding very well. And from that standpoint, we called that particular chipset ARTPEC, and that stands for Axis Real Time Picture Encoding Chip. And from 1996 to now, we’ve iterated in several generations of chipsets since then to improve the processing power at the edge. And right now we’re in our eighth generation chipset and it’s very exciting. And why does that all relate to AI? I’ll give you a little bit of a background in terms of why it relates to AI.
Quang Trinh:
And that really comes down to is if you think about AI, it’s a huge buzzword right now similar to the buzzwords of cloud and cyber security, a couple years ago. And right now, AI is a very transformative technology. It’s going to change the industry and every industry. Not only our security industry, but other industry that are involved. And from our standpoint as a manufacturer of edge devices, what we did with our chipset is a little bit of a differentiator compared to what other companies have because we actually control our own destiny. We design, we develop, we optimize.
Quang Trinh:
And about a couple of years ago starting with our sixth generation chipset, we moved from ASIC to a SOC type of architecture. And with that SOC type of architecture we were able to embed a GPU into our edge devices so that we can do some additional processing and offloading because we were dealing with video. There was a lot of things that our customers wanted with overlays and things like that. But also at the same time, in 2009, that was when we first brought out our software platform. It’s called ACAP. So that’s Axis Camera Application Platform.
Quang Trinh:
So similar to how you guys load apps on a phone, you can actually load apps and programs on the edge and run it and let it operate things that are just beyond what the device was doing from a video perspective. So that ACAP platform started in 2009. And that’s a very important piece because that allowed a lot of our developers and a lot of our partners to take advantage of the additional horsepower, the additional processing power at the edge to do some very unique things to enhance a lot of the solutions that they were developing at that time.
Quang Trinh:
So how does this all relate to AI? Well, the one thing about machine learning and deep learning under the AI umbrella is that it’s becoming in the forefront of a lot of different companies right now. We’re living in a digital age where everything is on the network, everything is digitized. And the data that’s being collected from all these different devices are becoming more and more very valuable.
Quang Trinh:
And from that standpoint, machine learning has always been very structured. It’s able to take information from database, whether you are an online retailer with a database of customers’ names, locations, zip codes, phone numbers, products, machine learning was able to put together predictions and trends and things like that off of very structured data.
Quang Trinh:
Now with the advancements of machine learning and their techniques and the evolution of that into deep learning and neural networks, a lot of the data that they’re able to process is actually unstructured data. And they do that really well. Why does that relate to us? Well, on the video side, we handle a lot of video and images. And video and images are what a lot of the AI community call unstructured data and they love that. They love to feed on that unstructured data so they can build models on things in terms of how they can increase the intelligence of their models and their AI solutions.
Quang Trinh:
So from that, there’s a huge synergy with the AI community and the community in the security industry, especially that these devices are on the network. They’re already digitizing the data, the video data. And this is a very huge inflection point that’s going to help fuel that fire to breed up more innovation within deep learning and machine learning, and neural networks to really elevate AI to the next level. And this is why there’s a lot of analysts out there that are predicting that AI is going to accelerate exponentially over the next couple years, if not the next decade. So this is very important when it comes to understanding how that plays into what we do here at Axis when it comes to edge devices and what we’re preparing them for the future.
Quang Trinh:
Now, the one thing is that at Axis we’re very forward-thinking, so we like to plan things five, 10 years ahead. And a lot of things that we do are in anticipation of a lot of these different technologies that might come into our industry and impact it in a very positive way. So with that, a couple of years ago when we had our release of our seventh generation ARTPEC-7 chipset, we actually collaborated with a global company that developed their own ASIC processor, their own tensor processing unit that handles deep learning, machine learning and AI type of techniques. And it was a very well done architecture. And we worked with them to actually implement that chipset, that particular chip as a co-chip within our ARTPEC-7 devices.
Quang Trinh:
And that was released a couple of years ago and that actually allowed a lot of our customers to evaluate and entertain a lot of the things that they were doing with AI. And going back to that ACAP platform that I told you about in terms of our application platform that we started in 2009, and that we evolved since then. The nice thing about this is that with AI it’s all about building models and seeing how much they can push that model down to the edge, and what they can do from a processing standpoint. And from there, seeing what they can do.
Quang Trinh:
And because we had our ACAP platform, it allowed for our customers to experiment with a lot of these different image classification models that they were building, packetizing it down to a small package and then running it on the edge on our devices, because we had the software framework for them to actually do that and test it. So with that, we’ve learned a lot of different feedbacks from our customers and our partners.
Quang Trinh:
And then a couple of months ago when we announced our ARTPEC-8 chipset and that particular SOC now has an embedded deep learning processing unit. And because of that, we’re trying to simplify that for our customers. Because when we released the ARTPEC-7 with that particular co-processor that handled the deep learning, it was only on a handful of products. And then our customers had to figure out which products had that particular chipset.
Quang Trinh:
Well, now, moving forward anything with an ARTPEC-8 chipset is going to be able to handle deep learning processing. Because that particular process or architecture that we’ve learned from is now embedded into our SOC design. And moving forward, that’s going to be part of our DNA. So we want to make sure that in order to accelerate AI, we want to make it easy for customers to adopt it.
Quang Trinh:
So whether they want to take advantage of it or not, having a hardware platform that already has it ready to go out of the box, is going to be very beneficial. It’s all about adoption. The more adoption we get with customers out there that want to look at different ways of utilizing AI, machine learning, deep learning as part of their overall solution is going to be very key for us. And that’s what is going to differentiate us between a lot of other companies.
Quang Trinh:
The other thing that’s differentiating us between a lot of other companies is our partnership. As you know, we work with not only Chooch, but we also work with a lot of companies across the stack. Not only from companies that work with us on the edge, but also companies on-prem and on the cloud. What we feel is that a lot of different and companies in the AI industry, they have a very common strategy. If you think about other industries, other industries you have companies that will try to sell you proprietary technology. And that’s really not what we’re here to do.
Quang Trinh:
We understand that AI’s going to be a very complex topic and technology. It’s going to require collaboration. And that type of strategy is what we call horizontal. And what we mean by horizontal is we don’t look at and differentiating our companies out there as competitors. They’re more as partnerships. So we will work with anybody that wants to work with us. And by doing that, we will be able to cultivate a solution for the customer because we know that it’s not going to be just one partner or one company that’s going to solve a customer’s problem. It’s going to take several companies coming together to actually collaborate and actually come to a conclusion or a solution that’s going to benefit the customer’s use case.
Quang Trinh:
And from that standpoint, we feel that that partnership is going to be very key. And in order to facilitate that partnership, what we’ve done is we’ve opened up our ACAP platform to open-source. And what we mean by open-source is now it’s available for anyone, people inside the security industry, people outside other industries to take a look at our platform and look at what they can do with AI. With not only cloud on-prem and edge, but play around with the whole ecosystem and that we believe is going to drive innovation.
Quang Trinh:
So with that open platform, that open nature of our development framework, we hope to have smooth integrations into a lot of solutions out there that are on-prem, that are also cloud-based. And we know that over the next couple of years, the processing that we’re going to be able to produce on the edge is going to accelerate and that’s going to be very exciting. So there’s going to be a lot of things where the solution might be edge and on-prem or edge and cloud, or a combination of all three.
Quang Trinh:
So really for us it’s all about that horizontal strategy and, and collaboration and open-source nature that’s really what’s going to drive the AI technology into our industry and beyond to other industries as well too. And from that a lot of things that that has been happening is a lot of our customers now are coming to us with some very unique use cases on a lot of different things that we’re working on with on to try to solve. And some of these use cases are really driven by the end customer and they’re not even public safety-related. They’re related to a lot of other different problems that they’re facing that are non-public safety or security-related. They’re operational. There’s a lot of different things that they’re doing in from a marketing standpoint.
Quang Trinh:
So one use case that I can talk about was very interesting is that a customer had some problems with monitoring emergency exits and they wanted to see when an emergency exit was blocked because they were getting fined by OSHA. If you block an emergency exit, you’re going to have issues with not only government fines, but then you might have issues with liability because a lot of things can happen during an emergency. And if you’re blocking an emergency exit, there can be some liabilities and some monetary compensation that’s a risk for a particular company.
Quang Trinh:
So what we did was we actually investigated that particular solution. And when we looked at it, we were like, “Okay, we can build a model and we can look at looking at AI and machine learning and deep learning and building a model to try to solve the problem or we can look at other possible solutions.” So the other solution we looked at was.
Quang Trinh:
Computer vision. So the one thing about computer vision and learning that a lot of people don’t understand is that with computer vision, there’s a lot of data that the sensor itself already sees on the edge before it gets into a machine learning model where it’s going to learn what it’s seeing. So from a computer vision standpoint, there’s a lot of things that you can get from the image sensor because the image sensor before, it looks at the data and puts it into a pretty picture that everybody here can see and relate to.
Quang Trinh:
There’s a lot of data sets that are there. So if a lot of people out there that are familiar with photo editing, the RGB spectrum with red, green, and blue. These are different things that computer vision can actually identify ahead of time. They can look at all the pixels on the sensor and figure out different hues, different tones and things like that. And they can actually segment foreground, background and things like that, but not go into the more deeper things.
Quang Trinh:
Because in order for them to classify, the more deeper things, you got to take all that computer vision data and send it to a machine learning model or deep learning model for them to understand what it is that they’re, they’re looking at. So if I’m just looking for a particular object, that’s a red color, then I can use some computer revision data to kind of look at, “Okay, what are the reds in the scene?” And then from there, extrapolate the pixels from that red. And then maybe encode that into an image and then send it over to like a machine learning model so that it can start to classify that particular object that’s red.
Quang Trinh:
So there’s a lot of different use cases. And the thing about computer vision is that computer vision is also very subjective, and it really depends on the sensor type and also the processor that’s handling the computer vision data. So not every camera out there is going to have the same computer vision data. So it comes down to the sensor sensitivity, the processing of that particular data before it becomes an actual image. So from there you want to have a good product or a good sensor that’s going to give you quality information. And from there, then you’ll be able to extrapolate the information that you need.
Quang Trinh:
So in this particular use case in terms of blocking an exit we felt that, “Okay, it might take a while for us to train a particular model to basically classify several different variety of objects” because we just don’t know. The customer doesn’t know what kind of objects that are going to be blocking that exit. So from that standpoint instead of going towards the AI machine learning route, we actually went towards a computer vision route. And what we did was we use an open CV, that’s an open-source computer vision framework that’s available out there. And we looked at a technology called entropy.
Quang Trinh:
So with entropy, it’s kind of like a histogram of a particular image. It gives you a mathematical value of the image. And what that allows us to do is a lot of these emergency exits are very sterile, very static. The cameras were positioned over the doors in a way where nothing shouldn’t be there. It had consistent lighting throughout the day. And what we did was by using open CV and entropy, we were able to capture an entropy value of that particular scene at it as a baseline.
Quang Trinh:
And then now anytime an object gets into the scene, that entropy value’s going to change, the histogram of that image is going to change. And because of that, we were able to set thresholds and sensitivities and alert on that because we don’t know what object’s going to be there, what size of the object. So based off of using this particular open CV computer vision technology on the edge, we were able to create an application that runs on the edge, that’s able to at least identify.
Quang Trinh:
And then from there, it sends a snapshot and alert to a particular VMS or video management system so somebody can visually verify if the emergency exit is blocked. But at the same time, we also tied that to an audio IP device where it created an automated message so that if somebody was putting something there to block the exit, it would alert them to please remove that object. So it was a very unique use case that involved a little bit … It started with conversations around AI and then we ended up using some computer vision technologies to actually solve the problem.
Quang Trinh:
Now, the one thing about computer vision and machine learning and AI and deep learning is that when you combine the two, it can be very powerful. And that’s where the second use case that I have to talk about briefly is a very unique use case. Because we’ve always been asked for the last couple of years, can we identify certain animals, certain objects? And with machine learning and deep learning, we’re starting to get to the point where we’re able to classify people and vehicles. But with animals, there’s so many different animals to choose from.
Quang Trinh:
So in this case, we actually worked on a project in terms of utilizing both edge, on-prem and cloud. So this is an example of using the full stack to solve this particular problem. So this particular analytic is to detect dogs in a particular training facility. So what we did was we actually wrote a particular code running it on a Jetson Nano board that’s an on-prem device connected to an Axis camera. It’s a pan, tilt, zoom camera, PTZ, that can actually move, zoom in. And we were trying to create an auto tracking solution where once it detected a dog, it would track that dog throughout the whole training facility so that it would always be in the field of view.
Quang Trinh:
So with that, this particular Axis camera did not have the processing power to run analytics or AI or machine learning at the edge. So we used a Jetson Nano board that can run a very simplified model for image classification. But at the same time, it wasn’t powerful enough to do the image classification that we needed. So we tied that to a cloud platform with YOLOv5 and a neural network that’s out there that already has a huge library of existing models for detecting dogs.
Quang Trinh:
So with that, we connected the dots. We connected this particular neural network that already has the models and the image and the data on identifying dogs. The Jetson Nano was the intermediate on-prem device that’s going to be able to process the data between the edge devices, to the cloud and back and forth. And then the other thing is that once we got the information from the camera sent up to the neural network and its identified that this is a dog, that information needs to be sent back. And then from there, we need to be able to code the PTZ to follow and track the coordinates of where that image was identified.
Quang Trinh:
So this particular solution is very unique. If you guys follow me on LinkedIn, I actually posted a video of this solution on my LinkedIn that you can take a look at. But those are hopefully two examples, some use cases that talks a little bit about computer vision that also talks about little bit about AI, machine learning and deep learning and how that all ties in within the whole stack. It’s a very technical topic and I can speak for hours on it, but I want to make sure that you guys are able to kind of get a high level overview of some of the unique things that can be done with AI, machine learning and deep learning and edge devices.
Quang Trinh:
And that’s where we work with Chooch a lot. Chooch is a great company when it comes to accelerating and building out models and they give us great feedback on the type of images that they would want to capture from. And there’s a lot of things around that’s accelerating the training around building models and Chooch has a really good solution on accelerating the of image classifications. And the thing that we try to advocate is a lot of people know that with machine learning, deep learning, it’s all about data. And they want as much data as possible, but you know what? It is. But then it’s also not because it’s also about the quality of the data.
Quang Trinh:
So in the technical world when it comes to programming, we have this thing where we call garbage in, garbage out. So if you send in a lot of bad data, you’re going to get a lot of bad data on the output. So really when it comes to the unstructured data that machine learning and deep learning loves to be able to learn from from images and video, it’s about the quality of the data. So if we can get a good quality video and images, you’re going to get very good quality output. So that’s very important to understand is that not everything is apples to apples.
Quang Trinh:
And going back to that computer vision aspect of the computer vision side of the sensor, it understands the intensity of color. Because computers, they don’t know what red, blue or what white is. They know it through values in terms of the RGB spectrum and the numbers we put into those particular intensity pixels. So with that, you have to worry that certain sensors they might not detect red as red. It might be a little bit off hue or off intensity. So this is very important for the machine learning world and the deep learning world is that getting good quality data? Probably trump’s the quantity of the data.
Quang Trinh:
Yeah, everybody wants more data to build better models, but you can still do really good outcomes with good quality data and small sets as well, too. But it’s all in the process of how you do that. And I know Chooch has been very good with the way they implement and build models. And this is why we like working with them. And we’re always open to feedback on what can be done on the edge. We know we’re not the end goal here. There’s a lot of things that can be done in collaboration. And that’s why that horizontal strategy is so key to us, but we also know that our edge done devices are becoming more and more powerful.
Quang Trinh:
It’s going to offer more information and more data for customers like Chooch and for our end customers. So we want to know what kind of information they want to get access and to help accelerate the AI technology as a whole. So with that, for us, our approach is always this ecosystem solutions approach. We’re going to need partnerships out there to help a lot of our customers’ problem and they have a lot of unique problems. And that’s the thing that we thrive on. We love the challenge. We love to be challenged because we feel that we can actually solve a lot of these problems with technology.
Quang Trinh:
So with that, I like to close my end of this segment to basically state that at Axis here, we have a model of innovating for a smarter and safer world. And we’ve for the last couple of decades we’ve been working on the safer side really well with getting everything on the network and really applying that to a lot of our diversified portfolio. On the smarter side, we’ve been doing that also in the last couple of decades with a lot of different solutions that are being run on the edge.
Quang Trinh:
But this AI technology, this paradigm is going to be very, very impactful not only in our industry, but in other industries. So we know that it’s going to be beyond just public safety and security. It’s going to go into operational efficiencies. It’s going to go into marketing. It’s going to go into other different realms where video is going to be very valuable, and these are going to be very valuable. And from that standpoint, we see very positive outlooks on how AI is going to impact our industry, how it’s going to impact us as a company.
Quang Trinh:
And with that, I’d like to thank you for your time and really thank you to Chooch and the rest of the people here that invited me to come here and speak on Axis’ behalf. So have a great day, guys. Take care.
