This $400 supercomputer might replace a $10,000 Nvidia product for many drone use cases

A team of wildlife biologists, researchers and drone experts might have just proven the worthiness of a $400 solution to a process that would otherwise cost $10,000. And while their use case is to use drones to protect endangered animals in Namibia, such as rhinos, the results of their research could save all sorts of drone companies thousands of dollars. Most of it is pegged around a $400 supercomputer from Nvidia called the Jetson Xavier NX.

Pair it with a Parrot Anafi drone, and the rhino-saving team suggested that they’ve devised a way to efficiently and cheaply use drones to seriously cut down on poaching. And even if you’re not using drones to reduce poaching, adoption of this $400 supercomputer tech via the Jetson Xavier NX could have bigger implications for the broader drone industry.

Jetson Xavier NX Developer Kit
The Nvidia Jetson NX

A wealth of information was laid out in a comprehensive study published this summer in PeerJ, an open access journal for life and environment.

A history of using drones to reduce poaching

Reducing poaching of endangered animals has long been a desired use case of drones, even in the early days of the technology. Even as far back as 2012, Google gave $5 million to the World Wildlife Fund to purchase drones to fly over parts of Africa and Asia in an attempt to help monitor and catch wildlife poachers.

In 2016, the now-defunct drone startup Airware used its platforms in Kenya to monitor wildlife in an effort to protect endangered rhinos. A year later, the Tanzania National Parks Authority (TANAPA) signed on to use drones as a form of anti-poaching surveillance, working with drone anti-poaching service Bathawk Recon to operate the drones.

Rhino poaching is a huge and ongoing problem. Rhino horn, which is used in some types of Asian traditional medicine, reportedly sells for around $65,000 per kg. Today, there are only about 5,500 black rhinos remaining on earth.

To combat it, conservationists have turned to a myriad of sensing technologies. Among common sensor tech that has been deployed:

  • Wearable tech: Collars and tags on the animals can work, but that tech often has poor longevity and presents risks for the animals. Some studies suggest collars reduce the animals’ fertility.
  • Ground-based camera-traps: Ground-based camera-traps are seen as effective in theory, but they need to be deployed in large numbers to make a meaningful impact, which can be expensive.
  • Small fixed-wing aircraft: This is actually the primary cause of mortality in wildlife biologists. And, it’s expensive.
  • Commercial satellites: They work for large animals like African elephants at the landscape scale, but their resolution is not yet sufficient to accurately count smaller species like black rhino and large antelope.

Given all those pitfalls, drones have become another potential, workable sensor. But even a drone on its own isn’t enough.

“Africa is too big to be simply launching small drones into the night sky with the hope of spotting rhinos or poachers by chance,” said University of Maryland professor Thomas Snitch, who builds analytical models to predict where the animals and poachers will be.

And because rhinos most often live in areas with poor wireless networks, drones can’t stream images back in real-time. So, if the goal was alerting authorities of poachers in real-time, that won’t exactly work, as the drone would have to land, and then researchers (or software) would have to comb through images and videos after the fact.

Aerial photo of a rhino via WildTrack.

New drone technology for today’s anti-poaching efforts

A team of researchers, technologist and veterinarians, included representatives from non-invasive monitoring company WildTrack, have sought out better tech solutions where drones are involved — but better drones.

Their tech of choice? A NVIDIA Jetson Xavier NX module onboard a Parrot Anafi drone. They built an AI using a YOLOv5l6 object-detection architecture, which they trained to identify a bounding box for one of five objects of interest in a video frame. The AI can account for differences in terrain, camera angles and lighting conditions.

Augmentation using tiling.

The NVIDIA Jetson Xavier NX is also ideal because the drone can still connect even when flying through relatively poor-quality wireless network areas, still able to deliver live notifications whenever the target species are spotted. That’s possible because the Jetson NX edge device performs inference on the drone during flight and only sends the detected video frame to the cloud rather than overloading the network and sending the entire live stream.

While not perfect, the researchers say they are confident that their model is pretty darn good. The AI correctly identified black rhinos 81 percent of the time. It also correctly identified giraffes 83 percent of the time, according to a research paper.

And perhaps most interestingly is that the Jetson Xavier NX Developer Kit has an MSRP of less than $400. Similar researched around building drone-based animal object detection models, published in 2019, used a Nvidia Quadro RTX 8000 GPU. That hardware costs $10,000.

The NVIDIA Quadro RTX 8000

What’s more, the newer (and cheaper) tech had a faster inference time at 30 fps, which is 15 times faster than the fastest of the two models being compared.

Where do drones for rhino research go from here?

“The model performs comparably to other published studies in terms of accuracy while having inference times that are an order of magnitude faster and running on cheaper hardware,” the researchers wrote. “We have also demonstrated a proof-of-concept edge implementation of a pipeline with a web app to guide potential real-world deployment. The combination of our model and implementation is ideal for low resource settings because a small edge device would be able to contain the lightweight YOLO model that can rapidly ingest and perform inference on captured imagery as the drone flies over large areas.”

Yet while the workable, low-cost solution feels promising, it’s not a surefire bet, as there are still plenty of other considerations to be made and challenges to account for.

For one, there’s still lots of ongoing discussion around actually flying drones around wildlife so as to not disrupt them — in particular debate around the appropriate altitude AGL to fly drones for conservation. Those factors include:

  • species sensitivity
  • noise generated by the specific aircraft
  • wind and air pressure variables
  • direction of approach

Flying lower versus higher has plenty of pros and cons. Flying higher means less animal disturbance and ability to cover more ground. Flying lower means higher-quality images and greater detection rates. For now, 30 to 40 m AGL is commonly accepted as an appropriate AGL.

Other questions like personnel training and operations need to be resolved before wildlife managers can implement such a system in practice.

Bringing the drones + Jetson Xavier tech to other applications

But it’s not just rhino research that could benefit. The study’s leaders said that, with small adjustments, the same basic system could be adapted to other types of drone use cases.

You might be able to get accurate population estimates from object detection models run on drone footage. Further work could also focus on how to integrate this type of pipeline into a whole system.

“We have shown that modern hardware and open-source software are capable of the task in an on-board edge device, the researchers said.

More details about the Jetson Xavier NX Developer Kit

The Jetson Xavier NX Developer Kit, with an MSRP of just $399, might be the solution. It’s designed for intelligent machine OEMs, start-ups and AI application developers who want to create breakthrough products — and an anti-poachng drone might just be one of them.

The NVIDIA® Jetson Xavier NX Developer Kit gives you supercomputer performance with a Jetson Xavier NX module. With the kit comes both the power-efficient, small form factor Jetson Xavier NX module and reference carrier board, plus AC power supply.

Jetson Xavier NX Developer Kit

Use it to build your own, multi-modal AI applications with the NVIDIA software stack in as little as 10 W, with what Nvidia claims is more than 10x the performance of its widely adopted predecessor, the Jetson TX2. It also offers cloud-native support to let you more easily develop and deploy AI software to edge devices. And of course, it’s supported by the entire NVIDIA software stack, including accelerated SDKs and other NVIDIA tools for application development and optimization.

Among its specs:

  • AI Perf: 21 TOPS
  • GPU: 384-core NVIDIA Volta™ GPU with 48 Tensor Cores
  • CPU: 6-core NVIDIA Carmel ARM®v8.2 64-bit CPU 6 MB L2 + 4 MB L3
  • Memory: 8 GB 128-bit LPDDR4x 59.7GB/s
  • Storage: microSD (Card not included)

Of course that low price tag has created higher demand, and it’s hard to actually come by — as it seems to be out of stock most everywhere. It’s currently being sold by a third-party on Amazon for a not-bad, but higher $539. It’s completely out of stock on Colorado-based electrical components website Arrow.

Grab your own Jetson Xavier NX Developer Kit now. And if you’re interested in digging far deeper into the rhino research and how the Jetson Xavier NX Developer Kit was used, check out the full research study here.

One Comment

  • Ed says:

    That is by no means a supercomputer which are compared in Petaflops, not Teraflops. It’s akin to comparing someone flying a drone to someone flying a top end fighter jet as if they are the same.

    The current 10th ranked supercomputer can do 61.61 petaflops. i.e. it is more than 2900 times more powerful than this.

    The #1 supercomputer is over 80,000 times as powerful as this little supposed supercomputer.

Leave a Reply