Contactless biometrics for infants
Babies are the people most likely to be missed by health records, and the hardest to identify. I lead a $6m initiative to develop the first reliable contactless infant biometric recognition algorithm, from data collection through model training, evaluation and deployment.
Three things make it hard:
- No data to start from. There are no public datasets of infant biometrics, especially from low- and middle-income countries.
- The subject keeps changing. Infants grow fast, so a model has to recognise the same child months later.
- Edge only. Capture and matching run fully contactless on a low-spec Android phone, with no cloud and only the phone camera.
Leading a team of two ML engineers, a backend engineer and, earlier in the programme, a data scientist, we built a longitudinal dataset of over 4 million images from a cohort of 13,000 infants in Ghana and Bangladesh, captured monthly for 12 months, and raised the true match rate at 1% false match rate from around 40% to around 90% over a three-month growth period.