Skip to Main content Skip to Navigation
Master Thesis

Reconnaissance d’activité humaine en natation : apports du deep learning pour le monitoring de l’entraînement basé centrale inertielle

Abstract : Objectives : The aim of this study was to develop an Inertial Measurement Unit (IMU) and deep learning based swim activity recognition system. It must be (1) to collect and labeled a large databse with intra et inter-subject variability, (2) to train a generalizable swim activity recognition model, (3) to use it for training monitoring, especially by computation of spatio-temporal performance and swimming efficiency metrics. Method : 35 recreational to national swimmers or triathletes were asked to perform a filmed and sacrum IMU equippped protocol including all swim styles, turn type and intensities. After manual labelling of recorded video in 8 activity class (rest, wallpush, underwater, turn and the fourth swim styles) a neural network based on bi-long short term memory is train on training set and evaluate on validation and testing set with a leave-one-subject-out method. A home-made post processing algorithm was also developped to extract from activity model prediction performance variables. Validation is achieve by video comparison. Results : Post-training accuracy was respectively 95,14, 94,36 and 93,65% on training, validation and testing set. Macro average F1-score by classe was 0,86 and weighted average 0,96. Activity prediction is faithful by real activity and many load and performance metrics could be computed, in particular lap time with a mean absolute error of 0,42s (n=952). Discussion and conclusion : Classification activity model show good generalization on unknow initial training data and with a larger number of activity class than in previous litterature. It enables, in compliance with high performance requirements, a non-invasive and automatic data collection in ecological conditions. This greatly facilitate swim monitoring for coaches. Validation of all the computed metrics is a futur way of work. As well as application and enrichment of the system with athletes and trainers.
Complete list of metadata
Contributor : Ufr Aps Rennes 2 Connect in order to contact the contributor
Submitted on : Tuesday, July 6, 2021 - 5:32:13 PM
Last modification on : Monday, October 11, 2021 - 2:24:03 PM
Long-term archiving on: : Thursday, October 7, 2021 - 7:11:33 PM


Files produced by the author(s)


Distributed under a Creative Commons Attribution - NonCommercial - NoDerivatives 4.0 International License


  • HAL Id : dumas-03279856, version 1


Antoine Bouvet. Reconnaissance d’activité humaine en natation : apports du deep learning pour le monitoring de l’entraînement basé centrale inertielle. Environnements Informatiques pour l'Apprentissage Humain. 2021. ⟨dumas-03279856⟩



Record views


Files downloads