Issue link: https://iconnect007.uberflip.com/i/1172746
OCTOBER 2019 I SMT007 MAGAZINE 27 before they arrive. Now, we have created a bur- glar alarm system for Africa's national parks that alerts rangers in near real-time of intruders coming into the park. Headquarters receives actionable evidence (i.e., transmitted images of humans) and can decide the level of response needed depending on if humans detected are heavily armed poachers versus herders trying to recover lost cattle. We've had great success in the Serengeti with one of our prototype units that worked for 1.5 years and detected 50+ intruders, including 30 poachers arrested in 15 separate incidents. Even several years ago, we knew it worked, and we've gone through four itera - tions since then, like any piece of technology. But the greatest breakthrough came when we gained the attention of Intel and were able to incorporate their vision processing unit (VPU) system-on-a-chip known as the Intel Movid - ius Myriad 2, which allows for deep neural networks (DNNs) to be run on the chip so that we can do inference on the edge. This is a huge advancement because most wild- life camera traps, or trail cameras, have not changed much from a technological perspec- tive. A trail camera I used in 1990 was one of the first ever built for commercial use. I tried it out to identify and photograph the elusive Javan rhino—the rarest mammal on earth— on the western tip of Java near the Krakatoa volcano. Until today, no wildlife camera had incorpo- rated AI. The value of having the inference on the edge is that 75–95% of the images recorded by camera traps have no value; they're trig- gered by the motions that pass infrared motion sensors, such as wind, branches, or a flying bird. Nevertheless, transmitting all of those empty or useless images cost you precious bat- tery life. And the two critical features you have to do if you're going to put a sensor in the field is to greatly extend battery life and address connectivity. We solved both of those problems by using AI to only send images of target species (i.e., humans) over our network. In the Serengeti, 95% of the images were of no interest, but we sent everything in our first try. By adding in the AI chip to permit interference on the edge, we're able to send a limited number of images that the rangers know are going to contain pictures of humans. This extends its battery life from about two months, which was what we had with our rechargeable lithium batter- ies, to now more than 1.5 years, which is a gamechanger. Johnson: Regarding battery life, is that more a design issue than cell technology? Dinerstein: Yes, because if you have hundreds of these units out, you don't want to have staff faced with the unnecessary time constraint of having to change batteries so frequently. That could possibly put them in harm's way or give away the locations of the hidden cameras. By using this chip, we've also been able to make our device low power. Almost all of the time that the TrailGuard AI unit is in the bush, it's in off mode. The system is only awakened for milliseconds when it's triggered by the motion sensor, which activates the whole system, and it wakes up in 100 milliseconds, which is extremely fast and critical. Next, it loads the AI module and does the inference on the edge. Then, it rates every image taken and sends the one that has the highest likelihood of being a human with a bounding box around the image. A text file with the probability that the subject in the image is a human based on a percent- Anna Bethke from Intel holding the TrailGuard AI PCB, showing Movidius Myriad 2 chip in red circle. (Source: Walden Kirsch/Intel Corporation)