As you may already know, one of the projects that our developers worked on (the expert AI team at NTR Lab) was a project for drones that can work indoors. NTR Lab created unique, embedded software for UAV autonomous indoor navigation.
The original object of the SLAM method was quite narrow. But more and more, as this technology becomes popular and more applicable on practice, we are moving swiftly towards a future we watched in movies.
There are many situations where people need a drone’s help, e.g., jobs that are dangerous for people to do and others that drones can assist people to perform faster and better. Among NTR’s development examples are indoor drone navigation for internal oil tank inspections, subway inspections, warehouses inventory, etc.
Before we succeeded we needed to solve many different problems for a variety of cases. We asked Head of R&D Department of NTR Lab Alexander Evdokimov what problems might arise when preparing a drone for work. He explained about the difficult conditions in which drones work and what tools can be used for drone navigation.
Drones indoors
There are three major impediments when using drones indoors.
- GPS does not work;
- compasses don’t always work correctly; and
- there is often little-to-no light.
One can try to use odometry-independent SLAM, e.g. Hector SLAM. However, the quality of positioning will be lower and it doesn’t work in a cylindrical oil tank, for example. It also requires additional equipment, such as lidar, which increases the weight of the drone thus lowering the payload weight.
Alternatively, you can use visual odometry and SLAM that supports it. This solution means adding one or two cameras and an extra computer. However, this again increases the drone’s weight.
When there is little-to-no lighting in the indoors environment the visual odometer doesn’t work well, if at all, so it is necessary to equip the drone with a fairly powerful light source. This again increases its weight.
Even with the additional lighting the visual odometer may still fail on a dark, glossy surface like we can see in an oil tank. A pulsed light source that is synchronized with the cameras significantly improves the situation, but does not solve it completely. Further improvement in stability can be achieved with a stabilized 2D / 3D gimbal for the odometer camera, once again increasing drone weight.
The continuing increase of weight also increases the size of the drone, which complicates its use with narrow openings, etc.
Our autonomous drone navigation solution largely solves the problems above for the specific case of oil and similar tanks for the purpose of structural inspections, but still does not allow navigation in dimly lit, large buildings with complex internal structure/geometry, like the ones encountered in rescue operations.
No effective solution to this challenge has been found. Currently, to obtain a quality image of the map in an unknown space it is necessary to increase the weight of the unmanned aerial vehicle. Our AI team believes the solution will result from the use of neural networks.
We plan to talk with other developers in our company who work on/with neutral networks and share their comments and thought in the near future.
Have you tried to navigate a drone? What do you think about drone navigation problems? Let us know in the comments below.
Share this article with your network.