更大功率, 更少的塔: AI 可能使飞机控制塔过时
这篇文章来自 nvidia.com。原始 url 是: https://blogs.nvidia.com/blog/2018/04/06/remote-airport-control-towers/
以下内容由机器翻译生成。如果您觉得可读性不好, 请阅读原文或 点击这里.
Airport control towers are an emblem of the aviation industry. A Canadian company wants to use its technology to make them a relic of the past.
Airport buffs may mourn the change. But Ontario-based Searidge Technologies believes its reasoning is, um, well-grounded.
It believes AI-powered video systems can better watch runways, taxiways and gate areas. By “seeing” airport operations through as many as 200 cameras, there’s no need for the sightline towers give air traffic controllers.
That doesn’t mean air traffic controllers are going away. The alternative Searidge proposes is a new concept made possible by remote towers. It’s not an easy idea to swallow for an industry that’s been reluctant to embrace change, and is sensitive to any perception safety is being compromised.
But the benefits are hard to deny, including reduced taxi and wait times, handling 15-30 percent more aircraft per hour and reducing the number of tarmac incidents.
“The industry is adapting, and often now puts air traffic controllers in regular buildings,” said Chris Thurow, head of research and development for Searidge. “It gives them a better view than they see out the tower.”
Originally a Radar Alternative
At first, Searidge focused on providing cheaper alternatives to expensive radar systems for tracking and identifying objects on airport runways and taxiways. The company’s earliest products used traditional computer vision algorithms that analyzed video feeds on CPUs. They met the demands on the system at the time, but that was more than a decade ago.
Since then, the resolution of video and need for real-time intelligence have both grown fast. CPUs can’t keep up with these resource-intensive features.
“Using GPU technology, we can offer this at a better price and with a significantly lower number of servers,” he said.
Searidge shifted to GPUs about two years ago. It also brought deep learning tools such as NVIDIA’s CUDA libraries, TensorRT deep learning inference optimizer, and the Caffe deep learning framework into the mix.
Then, as airports began to ask not only for coverage of runways and taxiways, but also tarmacs and gate areas, Searidge expanded the abilities of its technology.
The company started working on more advanced AI that could accommodate a wider range of business rules. This enabled it to detect a greater assortment of objects. It could even deduce when such objects might cause unexpected delays.
“We are still trying to find the limits of the technology,” Thurow said.
Trained with Pooled Airport Data
Searidge has been training its deep learning network on workstations running NVIDIA Quadro P6000 GPUs. The system constantly collects imagery from the airports it serves to expand its training base. Training typically takes five to seven days, so the company has recently begun training on the GPU-powered Google Cloud to speed the process.
The company deploys its technology on workstations running Quadro P6000 GPUs to do positioning of targets, classification and stitching of images in real time for 20 HD cameras. Once at a new airport, it annotates 24 hours of that facility’s normal operations and combines this with customer data from about three dozen airports in 20 countries — so its algorithms are always improving.
Searidge’s AI innovations are built on top of their “remote tower” platform. New control towers are no longer being built or renovated, Thurow said. Instead airports are moving air traffic control to ground facilities. They’re even considering off-site locations. With AI added to remote towers, they offer high levels of situational awareness and air traffic controller support.
In some cases, he said, smaller airports are considering joining forces, allowing a single remote tower to manage more than one facility.
The European Union’s first certified medium-size, multi-runway remote tower recently opened in Budapest, Hungary, using Searidge’s technology. All tower controllers have been trained on the system, which is initially being used for contingency operations, live training and as a backup system. By 2020, HungaroControl aims to operate a full-time remote tower at Budapest.
Eventually, Thurow believes further AI innovation will lead to a more fully functioning “AI assistant.” The assistant could help air traffic controllers by picking up things humans might miss, predicting situations and recognizing patterns.
“I expect AI assistants to come into play in the next five to ten years,” he said.