Design and implementation of a signal processing system for inertial sensors

Gil Martín, Manuel (2018). Design and implementation of a signal processing system for inertial sensors. Tesis (Master), E.T.S.I. Telecomunicación (UPM).

Descripción

Título: Design and implementation of a signal processing system for inertial sensors
Autor/es:
  • Gil Martín, Manuel
Director/es:
  • San Segundo Hernández, Rubén
Tipo de Documento: Tesis (Master)
Título del máster: Ingeniería de Telecomunicación
Fecha: 2018
Materias:
Palabras Clave Informales: Human Activity Recognition, Activities of Daily Living, PAMAP2 Data Set, Inertial Measurement Units, Fast Fourier Transform, Deep Learning, Convolutional Neuronal Networks, Recurrent Neural Networks, Rectified Linear Units, Leave One Subject Out, Octave, Theano, Keras.
Escuela: E.T.S.I. Telecomunicación (UPM)
Departamento: Ingeniería Electrónica
Licencias Creative Commons: Reconocimiento - Sin obra derivada - No comercial

Texto completo

[img]
Vista Previa
PDF (Document Portable Format) - Se necesita un visor de ficheros PDF, como GSview, Xpdf o Adobe Acrobat Reader
Descargar (5MB) | Vista Previa

Resumen

This master thesis deals with the design of a general deep learning architecture for recognizing human activities using inertial sensors. The whole system architecture is composed of several steps to recognise different human activities: data acquisition, signal processing, feature extraction and activity classification. Regarding data acquisition, we used the PAMAP2 database, which contains recordings of various physical daily activities such as sitting, running or ironing. These activities were performed by 9 subjects and the recordings were obtained from a heart rate monitor and three inertial measurement units placed in the dominant wrist, the chest, and the dominant ankle. For this work, only the accelerometer signals were used. This report includes a detailed analysis to characterize the recorded activities and to study the differences between activities and subjects. This analysis includes the distances between average feature vectors of the different subjects when a particular activity is performed and when a particular sensor is used. This analysis showed a high variability between subjects. During signal processing, the recordings from all subjects were conditioned and pre-processed. Then, these recordings were segmented into analysis windows. Finally, these windows were processed and transformed independently. In addition, in this work, we propose a new strategy for separating the gravity acceleration from accelerometer signals. This signal processing module was implemented using Octave. During feature extraction, we designed a deep learning architecture using convolutional neural networks (CNN) for feature extraction from human activities when using inertial sensors. Finally, a study of several recurrent neural networks (RNNs) was done for detecting the transitions between different movements. All the architectures proposed in this work were implemented using Keras (running on Theano). We conducted the design of the system by comparing different architectures and signal processing strategies, using the recognition accuracy as the evaluation metric. We used a Leave One Subject Out (LOSO) cross validation with subsets for training, validation and testing. In conclusion, promising results have been obtained with the FFT module and phase as the inputs to an architecture which combines several convolutional neuronal networks. We obtained a 4% absolute increase in performance comparing to the baseline. Moreover, some subjects’ analyses were done, observing that the higher the subject’s BMI, the better the system classifies the activity. It is important to highlight that the best results have been obtained when post-prediction filtering is applied, obtaining more robust decisions and reaching an accuracy of 97.86%. Some of the analyses done in this work have contributed to the preparation of a scientific paper in a high impact journal.

Más información

ID de Registro: 51572
Identificador DC: http://oa.upm.es/51572/
Identificador OAI: oai:oa.upm.es:51572
Depositado por: Biblioteca ETSI Telecomunicación
Depositado el: 10 Jul 2018 09:11
Ultima Modificación: 10 Jul 2018 09:11
  • GEO_UP4
  • Open Access
  • Open Access
  • Sherpa-Romeo
    Compruebe si la revista anglosajona en la que ha publicado un artículo permite también su publicación en abierto.
  • Dulcinea
    Compruebe si la revista española en la que ha publicado un artículo permite también su publicación en abierto.
  • Recolecta
  • InvestigaM
  • Observatorio I+D+i UPM
  • OpenCourseWare UPM