TrailGuard AI: Interview with Eric Dinerstein

Global Conservation is supporting TrailGuard AI, a revolutionary next-generation trail camera alert system that will help protect against human-wildlife conflict and illegal activities like poaching and logging. Created by the environmental organization RESOLVE, Intel, and software developer CVEDIA, TrailGuard is a very small camera embedded with a processing chip that, using the latest AI, automatically detects and alerts managers to people, vehicles, and wildlife.

Simulated examples of key wildlife species that could be detected and identified by TrailGuard AI.

TrailGuard AI revolutionizes trail cameras in several ways:

Its on-board artificial intelligence can be trained to recognize people, vehicles like logging trucks, and specific wildlife species, even in the dark.
TrailGuard AI is incredibly small – the head of the camera, along with the passive infrared sensor, is only about the size of a person’s index finger.
TrailGuard AI can operate for 1.5 years on a single rechargeable battery, rather than two months like most other trail cameras.
Using either cell, long-range radio, or satellite, TrailGuard AI offers multiple ways to transmit images in near-real time from the camera to the Internet and back to park in under 2 minutes.

Dr. Eric Dinerstein, a conservation biologist with RESOLVE and a Global Conservation Advisory Board Member, has been instrumental in developing this new technology. Over the past six years, Eric worked closely with lead inventor Steve Gulick—an expert radio engineer, computer scientist, and a field conservationist— and other tech partners to bring TrailGuard AI to fruition. We caught up with Eric to chat about TrailGuard.

Eric Dinerstein building the first 300 TrailGuard AI units by hand.

What’s unique about the TrailGuard AI?

Eric: TrailGuard AI is an AI-augmented or -enabled camera-based alert system. We really don't think of it as a wildlife camera trap; we think of it as a complete end-to-end monitoring system. First, a motion sensor is triggered, whether it's by an intruder, a logging truck, an elephant heading to raid croplands, or a tiger moving through the forest. Secondly, an AI algorithm is loaded in- camera that only selects images of what's important to you. It then transmits those important images directly to your phone or computer. And it does all of this quickly.

So to give an example, we just got detections of intruders in a park in East Africa. The intruder came through, triggered the motion sensor, and didn't see the camera, of course. The camera’s “brain” quickly woke up, grabbed the images, loaded the AI module, ran inference on those, wrote all of them to the SD card, picked the one that had the highest probability of being human, sent that to our comms network, sent that to the cell tower to a server in the states where it drew a bounding box around the human and sent it back to the headquarters in 50 seconds. It’s magic.

So that's basically what we've done: create magic through this technology, by blending the latest in AI from CVEDIA where they use models trained on synthetic data, and addressing critical issues of battery life and connectivity. Those were the two biggest problems and we've solved them.

What motivated you to innovate in this space?

Eric: About six years ago, I was at a conference hosted at Google that brought together leaders in conservation technology. After a number of presentations, an engineer at Google who was not part of the conservation technology space stood up and said, “You know, I think deploying large numbers of sensors in the field comes down to solving two problems: extending battery life and ensuring connectivity. You solve those and everything else will work.” I would add that price is another one.

Another part of what motivated me was the frustration of seeing, reading about, and hearing from colleagues who are doing extensive camera trapping that their cameras were constantly being stolen or vandalized. The first place that we deployed TrailGuard in Africa, there was an ongoing study that lost 40 percent of their cameras in the first year to theft and vandalism. With a typical research budget, it’s hard to deal with that kind of attrition. Besides, you’re losing valuable data. So we wondered, could we come up with a camera that's the size of my index finger and could easily be hidden?

How does the TrailGuard solve those problems?

Eric: It’s basically undetectable. The TrailGuard’s artificial intelligence is agnostic for the angle of the camera to the trail, so you could put our camera fifteen feet up and still detect the tiger or the person or the elephant and be out of the line of sight of a person walking on the trail. So I guess you could say that we have reinvented the “trailcam” as a “tree cam”. And the next version will be even smaller -- the size of a seedpod, and it has everything. The key realization was that we could move the batteries and communications unit out of the camera so the camera part could be concealed more easily.

Additionally -- a lot of traditional camera traps use a visible infrared illuminator, which makes the cameras tough to hide. The TrailGuard’s illuminator is set at a wavelength of nine hundred fifty nanometers, which is above the human threshold. Humans can't see it. But what’s even more exciting is that in the near future, we should be able to do away with the infrared illuminator. The newest AI software is clever enough to take a picture in near-total darkness and enhance it to make it look like daytime. We have this amazing new way to see in the dark.

Also, you don't want to be changing batteries every two months or so. That’s where the AI comes in -- it greatly prolongs battery life by only transmitting what the computer model thinks is a human or a tiger or an elephant. Often, in a field situation, from 75-95% of the images taken are of no value. A much smaller percentage of images are of actual human intruders or species of interest. If the camera sends all of them, it’s chewing up valuable battery life. And then if it’s an expensive satellite connection, it would be completely unaffordable. With the AI filtering the images, you drastically cut the number of images you’re transmitting.

TrailGuard AI units are nearly impossible to spot once properly camouflaged (left).

Besides, the device is miserly on power consumption. When it's not triggered, it's in off mode. It spends most of its life in a deep sleep. That way you don't have to go and change the batteries, which is a major task if you've got hundreds of cameras out there or if it's a dangerous place and the rangers are going to be in harm's way while changing them, or you're giving away the location of the camera because you're going back and forth so much.

The third thing is connectivity. The cell-based cameras work out of the box. It's fantastic -- if you just plug it in the right way, it automatically makes a connection with a local cell provider and transmits images. But one of the things that we've discovered in the last four months of starting to roll out TrailGuard in seven parks is that a lot of places where people think they have good cell connectivity, they really don't. They'll put the camera out one day and they can transmit images, but on the second or third day, there is no signal from that site. It's this intermittent signal that you can't control. So you need to look for other ways of transmitting images.

For that reason, we've adapted a technique called LoRa, or long-range radio, that's lower power and can transmit images over long distances. But what's emerging now is satellite connectivity. And that's the really exciting advance. We have just gained access to a low-cost satellite modem through our partner, Inmarsat. And the airtime provider, our other partner, Galaxy1, is offering us a data transmission package that makes this satellite-based alert system affordable for everyone. Imagine, now we can pair each camera with its own dedicated satellite modem. With this advance we should be able to transmit image alerts from even the most dense rainforests or remote islands.

TrailGuard AI is so desperately needed and there's nothing like it yet. If something like TrailGuard AI already existed, there would be millions of them out there. That tells us that we're on the right track, filling an unoccupied niche.

How is TrailGuard AI’s artificial intelligence different?

Eric: Typically, how computer scientists have been training neural networks is, they require ten thousand or more images of whatever species they want to detect. This large number was necessary to account for the variation in light conditions, background, distance from the camera, objects like trees or rock occluding the view of the object, etc. And then they'll start training the model using those images and improve it as they go. That process, from gathering images to training the model, is really time consuming. For rare species, it might be hard to obtain enough images, ideally taken from camera traps, to train the model. And even 10,000 images might be rather biased. With the old AI models, if you moved it somewhere else, you got lots of error. And the reason was that the model was probably training on objects in the background and not just the object of interest, which made it site-specific.

What CVEDIA does -- it's genius -- is they take a page out of Pixar or Disney’s book and they have artists create 3-D renditions of whatever object it is that you want to detect, whether it's a logging truck full or empty, or a tiger, bear, or human. And they animate that figure, that object. They expose it to all different lighting conditions and different angles and change backgrounds, so that the AI can recognize the object regardless of the light, the angle, or even whether there’s a big rock in front of it. They use that animation, that “synthetic data”, to create the model, and only use pictures from the camera traps to validate the model. That’s a much better way to do it, because it’s much faster, much cheaper, and more accurate.

Bear animations developed by CVEDIA are used to train the AI.

After training using animations, the AI can reliably identify real bears in a variety of settings and poses.

What are the biggest remaining challenges for the development of conservation technology?

Eric: There are a few really important things in conservation technology. First of all, it all has to move to the edge, meaning the AI algorithms need to be small enough to run on a computer vision chip in the camera rather than requiring that the data be transmitted to the cloud. It’s like autonomous driving. It’s impossible to have a self-driving car that takes whatever the camera is seeing, sends that to the cloud, where a centralized computer does the analysis and gives the answer back to the car. It takes too long, and you only have a split second to hit the brakes or swerve around the deer or the fallen tree or whatever. You need to have the capacity to run the inference on the edge.

TrailGuard AI can also be programmed to identify and transmit images of logging trucks.

The other issue for us in conservation is that computer vision chips have to be very low power. There’s a lot of computer vision technology out there, but it's very energy demanding. They consume a lot of power and they get really hot, too. Movidius, which was bought by Intel, made the lowest power chips available, the Myriad 2. Not only are they low power consumption, but they offer multiple modes of operation. They can be put in sleep or off mode, which is what we use. When the camera is woken up by the motion sensor, the computer vision chip is engaged for a few milliseconds, and then it goes back to sleep.

Finally, a major problem is connectivity. Most of the places we care about are remote or semi-remote, meaning they don’t have adequate or any cell connectivity. Cell providers just aren’t pointing their antennas towards the direction of the parks, but rather to the villages. That’s always going to be a problem.

But what if we could get that to the point where it's cents on the dollar for transmitting images? Then all of a sudden the world opens up.

Where do you see conservation technology going in the future?

Eric: I have a grand vision for conservation technology. Greg Asner – in my view, the top environmental scientist working today – is about to launch a program that is going to change the face of biology. His team does airborne spectroscopy using light aircraft, but now he finally has the opportunity to put those spectrometers into Space. With that, it will be possible to have a new vegetation map of the earth and assessment of coral reef health every few weeks. That's going to give us something that we never had before – we’ll be able to see biodiversity from space and monitor it.

But what about what the satellites and the spectrometers can't see from space? For example, on land, what about the animals living in those forests? Are they empty forests? Do they have the full complement of large carnivores and herbivores? That's where TrailGuard AI could come in. That's my grand vision: spectroscopy from space and TrailGuard AI on the ground. I want every square kilometer of the Global Safety Net to have a sensor that provides us data in near real-time on the critical species we need to monitor. In ten years or less, this could be reality. It can’t come soon enough to protect wildlands.