InSight: Enabling NLOS Classification, Error Correction, and Anchor Selection on Resource-Constrained UWB Devices

The accuracy of ultra-wideband (UWB) ranging is severely
affected when the direct path between devices is partly or
fully occluded, i.e., in non-line-of-sight (NLOS) conditions.
To detect and correct erroneous ranging measurements,
many solutions based on machine learning models have been
proposed, but they are usually deployed on edge devices
rather than on the UWB device itself. In fact, existing works
often focus on maximizing the NLOS classification accuracy and error correction performance, which results in large
and computationally-complex models that cannot be run
on UWB tags with limited processing power and memory.
Whilst convenient, off-loading NLOS classification and error correction tasks to an edge device severely affects, among
others, the scalability, privacy, and responsiveness of UWB based localization systems, as tags need to actively exchange
data with a third party and wait for its response, which may
be delayed due to heavy load or unreliable communication.
In this paper, we present InSight: a framework that enables
the deployment of NLOS classification and error correction
models directly on resource-constrained UWB devices.
InSight allows to train and generate such models according
to specific requirements (e.g., on memory usage and on
runtime), and to shed light on how to reduce the model size
and runtime without degrading the classification accuracy
and error correction performance. The selected models are
then seamlessly integrated into a NLOS engine running on
the device alongside existing applications and supporting
any localization service. With InSight, we can perform
NLOS classification and error correction directly on an
UWB tag in 0.6 ms, and with as little as 8 B of RAM and
19 kB of flash memory – while retaining a classification
accuracy of up to 86% and reducing the 90th-percentile
ranging error by more than 1 m. We further show how a
localization service can leverage InSight to select only anchors in direct line-of-sight and to correct erroneous NLOS
ranging measurements, which improves the 90th-percentile
localization error by up to 1.6 m on our 120 m2 testbed.