Bee: Learning sound recognition for camera-free security
| Reference number | |
| Coordinator | Minut AB |
| Funding from Vinnova | SEK 500 000 |
| Project duration | May 2016 - January 2017 |
| Status | Completed |
Important results from the project
The purpose of the project is to develop algorithms for learning sound recognition. A first application is surveillance in areas where privacy is expected. The goal has been reached and a system that can learn to recognise an arbitrary event using its acoustic signature has been implemented. The system comprises algorithms on a listening device which analyses the sound and sends extracted features to a server where machine learning algorithms are used to match the sound signature to a known event in an ever growing database. A large field-test with prototypes is ongoing.
Expected long term effects
The product is not ready for commercial applications and a lot of work remains before a possible introduction to the market. The breakthrough of this project is that the algorithms that form the foundation and makes it possible for the system to recognise new events has been shown to work in practice. If the system is commercialised and lives up to expectations, large opportunities to improve productivity and quality are opened in areas where privacy has to be balanced with a need for surveillance. Many applications are conceivable, for example in elderly care or in the insurance industry.
Approach and implementation
The system has been developed on a PC made to simulate the acoustic environment, the product and the server. Freely available sound libraries and ambient sound from sources such as Youtube have been used for training purposes. Parts of the software have then been implemented on a server and on prototypes that collect sound data in homes. Parts of the algorithms have been ported to existing hardware. Most of the computation have been done server-side which has enabled large scale field trials despite limitations of existing hardware.