Signal processing and algorithms for smart and wearable devices
A sensor gives you a signal. Turning it into a number people can rely on takes mathematics. We develop the signal processing and algorithms behind health, fitness, veterinary and wearable devices, and implement them on the device itself.
Who it is for
- Device makers whose product measures movement, heart rate, force or other signals and needs reliable numbers from them.
- Teams with a working prototype whose algorithm is not accurate enough, too slow or too power-hungry.
- Companies that want to add a new metric to an existing device without new hardware.
What we do
- Signal processing: filtering, noise and artefact removal, feature extraction.
- Sensor fusion: combining accelerometer, gyroscope, heart-rate and other sensors.
- Activity and event detection: recognising movements, steps, impacts and other patterns.
- On-device algorithms: implementations that fit the memory, processor and battery of a small device.
- Test methods: recordings, reference measurements and replay tools that make accuracy measurable.
- Apps: device, mobile and watch apps that show the results to the user.
An example we can show
Our own sport watch apps for Nordic walking count pole plants from the wrist accelerometer alone, with no extra sensors. On walks of 100 hand-counted plants, the first detector counted 168 and 116; the final one counted 104 and 106. The algorithm is about 300 lines of code and runs on the watch, offline.
How we work
- Record. We collect real recordings and reference measurements, not just lab data.
- Model and replay. We develop the algorithm offline and test every change against all recordings.
- Implement and verify. We port it to the device, measure accuracy, speed and power use, and document the limits honestly.
Frequently asked questions
Can you improve the accuracy of an existing device without new hardware?
Often, yes. Many accuracy problems come from the algorithm, not the sensor. Better filtering, timing rules and pattern recognition can extract much more from the same signal.
Do you work with raw sensor data?
Yes. Raw recordings are where we start: we look at the signal before choosing a method.
Which platforms do you work on?
Embedded firmware on microcontrollers, sport watches, mobile apps and back-end analysis. The algorithm is designed for the platform it will run on.
How do you prove that an algorithm works?
With recordings and reference counts made for that purpose, a replay tool that tests every version against the same data, and a written summary of accuracy and known limitations.
Have a signal that should become a reliable number? Write to info@kuorelab.fi.