-
Notifications
You must be signed in to change notification settings - Fork 2
3 Known Issues
On this page issues of the system are noted, which may be fixed in future (sub)-versions.
The position estimation is currently by brute-forcing a fixed grid. This works reasonably well for 2D inputs on approximately standard-sized tables. However, for especially large tables and many observed frequencies, this might have a notable impact on the pipeline's performance. Reference system use an non-linear optimization approach for the determination of the position based on the TDOAs, which seems to converge within several microseconds. Replacing the grid approach will make the processing pipeline fast enough for arbitrary realistic use-cases.
Currently, the accumulated FFT signals are smoothed in the temporal domain. For a high smoothing sample count or large signals, this takes at least a notable amount of time. Replacing this filtering by performing the convolution in frequency domain will increase the overall performance.
In future, the sound emitter app should be replaced by a web service, which should establish a data-link to the smartphones and stream the audio data. As a reference, the sine signal producing code of the smartphone app may be used.
So far, we tested our system extensively with Nexus 4 and Nexus 5 Devices. Although the Nexus 4 produced nice results with different settings, the Nexus 5 device seems to produce a more or less random offset to the device. This may be partly caused by the speaker characteristic and partly by the different position of the speaker.
We used relatively cheap microphones to do the recording. One would expect, that more high-end microphones will show an even better frequency response. However, with more expensive microphones, one sacrifices the affordability for casual users.