Gait recognition is a critical area of research in biometric identification, which has traditionally required complex instrumentation and systems for detection and analysis. However, wearable devices, such as smart glasses for gait recognition, prove the advantages of being less complex and more user-friendly than traditional vision-based systems. In this paper, we present our approach to gait recognition, which involves using smart glasses for gait data acquisition and deep learning as universal feature extractors. Our study aims to identify individuals based on their walking patterns using smart glasses’ accelerometer, gyro- scope, and magnetometer data. The findings suggest that smart glasses with embedded sensors can acquire the required data for training and testing machine learning algorithms, making detecting and recognizing gait in different scenarios possible. Essentially, the experimental findings also suggest that our approach leads to classification accuracy for multi-class (8 individuals classification) with an average accuracy of 93% and male/female classification with an average accuracy of 97%. Furthermore, the study addresses the challenge of maintaining the model’s performance on previously recognized and classified gaits after learning new ones without being retrained on their data. This challenge was addressed by adopting rehearsal method in continual learning to our model.
Il riconoscimento dell’andatura è un’area critica della ricerca nell’identificazione biometrica, che ha tradizionalmente richiesto strumentazione e sistemi complessi per il rilevamento e l’analisi. Tuttavia, i dispositivi indossabili, come gli occhiali intelligenti per il riconoscimento dell’andatura, dimostrano i vantaggi di essere meno complessi e più intuitivi rispetto ai tradizionali sistemi basati sulla visione. In questo documento, presentiamo il nostro approccio al riconoscimento dell’andatura, che prevede l’utilizzo di occhiali intelligenti per l’acquisizione dei dati sull’andatura e il deep learning come estrattori di caratteristiche universali. Il nostro studio mira a identificare gli individui in base ai loro schemi di deambulazione utilizzando i dati dell’accelerometro, del giroscopio e del magnetometro degli occhiali intelligenti. I risultati suggeriscono che gli occhiali intelligenti con sensori incorporati possono acquisire i dati richiesti per l’addestramento e il test degli algoritmi di apprendimento automatico, ren- dendo possibile il rilevamento e il riconoscimento dell’andatura in diversi scenari. In sostanza, i risultati sperimentali suggeriscono anche che il nostro approccio porta all’accuratezza della classificazione per multi-classe (classificazione di 8 individui) con un’accuratezza media del 93% e classificazione maschio/femmina con un’accuratezza media del 97%. Inoltre, lo studio affronta la sfida di mantenere le prestazioni del modello su andature precedentemente riconosciute e classificate dopo averne apprese di nuove senza essere riqualificato sui loro dati. Questa sfida è stata affrontata adottando il metodo di prova nell’apprendimento continuo del nostro modello.
Metodi di apprendimento continuo per il riconoscimento dell'andatura
OWUSU, THEOPHILUS YAW
2022/2023
Abstract
Gait recognition is a critical area of research in biometric identification, which has traditionally required complex instrumentation and systems for detection and analysis. However, wearable devices, such as smart glasses for gait recognition, prove the advantages of being less complex and more user-friendly than traditional vision-based systems. In this paper, we present our approach to gait recognition, which involves using smart glasses for gait data acquisition and deep learning as universal feature extractors. Our study aims to identify individuals based on their walking patterns using smart glasses’ accelerometer, gyro- scope, and magnetometer data. The findings suggest that smart glasses with embedded sensors can acquire the required data for training and testing machine learning algorithms, making detecting and recognizing gait in different scenarios possible. Essentially, the experimental findings also suggest that our approach leads to classification accuracy for multi-class (8 individuals classification) with an average accuracy of 93% and male/female classification with an average accuracy of 97%. Furthermore, the study addresses the challenge of maintaining the model’s performance on previously recognized and classified gaits after learning new ones without being retrained on their data. This challenge was addressed by adopting rehearsal method in continual learning to our model.File | Dimensione | Formato | |
---|---|---|---|
Owusu_Theophilus_Yaw.pdf
accesso aperto
Dimensione
5.25 MB
Formato
Adobe PDF
|
5.25 MB | Adobe PDF | Visualizza/Apri |
The text of this website © Università degli studi di Padova. Full Text are published under a non-exclusive license. Metadata are under a CC0 License
https://hdl.handle.net/20.500.12608/46150