A look back at Antoine Deleforge's participation at the IROS conference
From 1 to 5 October, Antoine Deleforge and Martin Strauss took part in the IROS conference in Madrid to present their paper entitled "Dregon: Dataset and Methods for UAV-embedded Sound Source Localization"
In this paper, Antoine Deleforge, Martin Strauss, Pol Mordel and Victor Miguet presenta new and original database made up of sound recordings captured by a drone in flight equipped with an eight-microphone antenna.
An application for rescue
In an emergency situation, these drones fitted with microphones could help rescuers to localise calls for help in hard-to-reach areas or when a lack of visibility prevents the use of embedded cameras.
The scientific challenges are those of sound localisation in noise conditions that are both extreme (that produced by the drone engines and the wind) and dynamic (the drone being in constant - sometimes rapid - movement). The biggest challenge in producing this database was the annotation of the audio recordings, in such a way as to be able to evaluate the performances of sound localisation methods. With the help of a sophisticated motion capture system (Vicon Tracking System), the position of the drone and that of a source present in the room could be annotated every 20 minutes with an accuracy in the order of a few centimetres.
An open database
The two researchers have made this database accessible to all via a dedicated website in the aim of advancing research in the fields of audio for drones and rescue. Their paper also presents the results obtained on the database with classic methods of source localisation. It has been possible to obtain a very accurate localisation in the order of a few degrees when the source emits a loud sound at high frequencies; however the localisation of human voices in such noise conditions remains a challenge today.
Following this publication, Antoine Deleforge is now coordinating the organisation of the IEEE's Signal Processing Cup 2019. From mid-November 2018 onwards, universities from around the world will be invited to test - in teams - sound localisation methods on these data. The three best teams will be selected by a jury and will be invited to the final, which will take place during the ICASSP 2019 conference in Brighton (UK) from 12 to 17 May.