Below is a translated article about our sound event recognition sensor SERVAL. To speed up development we open-sourced it, resulting in a great collaboration with the IoT Sensemakers AMS.
Enjoy the read!
How does the Marineterrein sound?
10 AUGUST 2021
Noise pollution is a big problem in the city. But to tackle it, you first need to know what exactly causes the noise. At the Marineterrein Amsterdam, tests are now being conducted with a sensor that can classify sounds and thus pinpoint the source. And this with technology that originated in the jungle to stop poachers.
It all started with a gunshot in the jungle of Laos, followed by the sound of a boat sailing away. It made Jan Kees Schakel of Sensing Clues realise that it is easy for poachers to make off with their 'loot' under cover of night. But he also realised that the sound the poachers produce is the way to stop them.
Sound sensor
Humans are noisy creatures', says Jan Kees. Basically any sound we make - talking, driving a vehicle and certainly gunshots - carries far and can therefore be picked up well by a sensor. A big advantage over cameras, which are limited by what their lens can 'see'. A sound sensor can help conservationists map out what and where is happening at any given time. After all, if voices can be heard in the jungle in the middle of the night, you can be reasonably sure that something is wrong.'
Complex
That sounds good, but it is easier said than done. Until about five years ago, only the number of decibels of a sound could be detected, but the sound could not be classified. In other words, a sensor could indicate that a loud sound was being produced somewhere, but not whether it came from a slamming car door or an elephant. And it is extremely difficult to be able to make that distinction,' says Jan Kees. You need an enormous database of sounds which then serve as a frame of reference for a complex algorithm to classify sounds'.
Practical challenges
'But since 2016, artificial intelligence and machine learning has taken off,' he continues. 'The technology is getting better and better and we can now recognise multiple sounds with one algorithm. But it is still important to have a large database of sounds and there are also practical challenges. There has to be power, a good internet connection and the hardware around the sensor has to be able to withstand the elements. To get everything working optimally we need to do a lot of testing.
Test location: urban jungle
Since testing sound sensors in the jungle is expensive and complicated, Jan Kees decided to do it closer to home. In the urban jungle to be precise. The Marineterrein has recently been equipped with a sound sensor that maps out city noise and noise pollution. Jan Kees joined forces with Sensemakers AMS, an old friend of ours who has been active at the Marineterrein for years, measuring and interpreting the water quality in the inner harbour, for example. With their combined knowledge, they developed a test set-up that will provide insight into the sound of the Marineterrein, and the extent to which there is nuisance.
The sensor of Sensing Clues and the IoT Sensemakers AMS
Training algorithms
With this project, we can gain a wealth of experience in training our algorithms', says Jan Kees. That is useful for conservationists, because in the jungle we look for the same kind of sounds as here. Voices, laughter, the sound of car engines and doors slamming. But with what we do here now, we can also tackle nuisances in the city. Noise pollution can reduce residents' enjoyment of living and even cause stress. But it is often difficult for people to say exactly what is bothering them. By classifying urban noise, we can pinpoint the exact time and source of various noises and use this information to tackle the problem in a targeted way.
Alarming noises
According to Jan Kees, this mainly concerns intrusive, loud noises. As a resident, you often no longer hear overflying aircraft or trams, you get used to it. But sudden, loud noises trigger us. It is still deeply ingrained in our cognition to be alarmed by such sounds, as a reaction to possible danger. Think of accelerating scooters and alarms, but also of bicycle bells. By classifying these types of sounds, we can pinpoint exactly where the nuisance comes from. It is important that we do this while preserving privacy. The sound is processed on the sensor and not stored. Only the sound labels, such as 'moped' or 'scooter alarm', are stored. So only labels remain, and no sound.'
Jungletech also for a liveable city
With the knowledge gained from this test at the Marineterrein, this technology can be used on a larger scale in the city in the future. In squares, in entertainment areas and even in individual cafés, to make visitors aware of the potential nuisance they are causing. In this way, Jan Kees hopes that his 'jungle technology' can contribute to a liveable city. But of course we eventually want to bring the lessons we learn here back to the jungle. That will require some further development: in the bush there are no power outlets or wifi, so we have to find good and cheap solutions for that. The most important thing is that the technology remains affordable for conservationists, so that they can always stay a step ahead of poachers.
Sensing Clues
Sensing Clues is a non-profit foundation and largely relies on volunteers. You can support Jan Kees' fight against poachers by making a donation. Do you have the knowledge and skills to contribute to this project? Then get in touch with Sensemakers AMS. Electronics engineers and people who like to get started with data analysis and visualisation are particularly welcome.
Text: Sjoerd Ponstein
Translated by Deepl.com