Skip to Main content Skip to Navigation

Sequence metric learning : Application to human activity recognition

Abstract : This thesis has been realized thanks to a cifre between Orange labs Grenoble and the LIRIS at Lyon. This thesis aims at proposing new neural network approaches to retrieve the habitual behaviors of fragile people in order to provide them with a monitoring at home while respecting their privacy and avoiding stigmatization. In this perspective, we concentrate on the exploitation of wearable motion sensor data (accelerometer, gyrometer, magnetometer, barometer etc.) which are nowadays easily embedded into smartphones and smartwatches. In a first contribution, we propose to employ few-shot learning with an architecture called matching network to learn a personalized and flexible activity recognition model. This model learn to recognize a new class from just one or few new samples since it matches rather than classify. Therefore this model allows to better handle the large variety of activities one can do in one day while alleviating the burden of data labeling. In a second part, we advocate for a change of perspectives by proposing to retrieve recurrent unlabeled activity patterns called routines instead of precise activities. We propose a formalization of the concept of routine with the notion of almost-periodic functions which prompts us to employ sequence metric learning. We propose a neural network architecture based on robust sequence representation learning with a Sequence-to-Sequence model and metric learning with a siamese network. No activity labels are used to train the model by setting up an equivalence constraint with the data timestamps. We propose to identify the routines with a spectral clustering and to evaluate the whole routine retrieval process with information-theoretic clustering scores. The last contribution of this thesis is a new neural network model for sequence metric learning called Coupled Gated Recurrent Unit. This model has been conceived by taking inspiration from the dynamical system theory and notably the concept of synchronization. We propose to improve the siamese gating recurrent unit architecture by implementing a coupling which should allow it to better process the hard samples. We finally experiment this architecture to recognize activities and retrieve routines.
Document type :
Complete list of metadata
Contributor : Abes Star :  Contact
Submitted on : Thursday, October 28, 2021 - 12:13:12 PM
Last modification on : Friday, October 29, 2021 - 3:58:19 AM


Version validated by the jury (STAR)


  • HAL Id : tel-03407186, version 1


Paul Compagnon. Sequence metric learning : Application to human activity recognition. Artificial Intelligence [cs.AI]. Université de Lyon, 2021. English. ⟨NNT : 2021LYSEI033⟩. ⟨tel-03407186⟩



Les métriques sont temporairement indisponibles