Filtragem de movimento em vídeos de captura biométrica neonatal

Abstract

Neonatal biometric identification is a critical field for hospital security, especially given the risks of newborn mix-ups and abductions. In this context, fingerprint recognition represents a promising yet challenging approach, due to the physical characteristics of newborns, such as small fingers, sensitive skin, and involuntary movement during data collection. To overcome these limitations, the use of videos of the finger in contact with the sensor is proposed, instead of single images, allowing multiple frames of the fingerprint to be extracted within the same session. However, this approach introduces an additional challenge: the newborn’s movement introduces abrupt variations between frames, undermining the fingerprint mosaicking process. This work evaluates frequency-domain filtering methods to attenuate these variations, with the objective of improving the quality of capture videos and increasing the biometric recognition rate. To this end, two approaches based on the Fourier transform combined with low-pass filters were implemented. The evaluation was conducted using image similarity metrics to assess motion filtering and recognition metrics to analyze the performance of the filtered videos in biometric identification systems. The results indicated that all tested configurations effectively reduced temporal variations in the videos. However, in the biometric recognition evaluation, only one of the tested configurations surpassed the performance of the unprocessed video. This result demonstrates that a balanced filtering value capable of improving the recognition process exists, but finding this balance proves to be a significant challenge. It is concluded that the proposed approach shows viability for improving neonatal biometric recognition, but requires the development of more systematic methods for the calibration of filtering parameters.