Prediction of football actions and identification of optimal sensor placements using a semi-supervised learning approach.
Authors
Larsen, Aske Gye ; Papi, Giovanni
Term
4. semester
Education
Publication year
2023
Submitted on
2023-06-01
Abstract
This study combines inertial measurement units (IMUs)—small wearable motion sensors—with Human Activity Recognition to automatically identify common football actions and to find the most informative sensor placements. The work had two parts. First, football actions such as pass, dribble, first touch, and positioning were predicted with a bidirectional Long Short-Term Memory neural network (LSTM), which is well suited to time-series data. Second, head scans (head movements that can occur at the same time as other actions) were predicted separately with a deep neural network (DNN). Seventeen players (14 male, 3 female) took part. Before modeling, the data were split 50/50 into labeled and unlabeled sets; the labeled portion was then split 80/20 into training and testing. All data were normalized and class-balanced using ADASYN. A total of 5,250 statistical time-domain features were computed in 200 ms sliding windows with 50% overlap and then reduced by Principal Component Analysis (PCA) while retaining >95% of the variance. A semi-supervised, uncertainty-aware pseudo-labeling method was used to reduce manual labeling time. The LSTM achieved solid performance on football actions, with a cross-validation score of 0.74 and an F1-score of 0.74. The DNN for head scans performed slightly better overall, mainly due to the smaller number of classes, with a cross-validation score of 0.79 and an F1-score of 0.78. The most informative placement for the LSTM was the right-calf sensor (F1-score 0.65), while for the DNN it was the head-mounted sensor (F1-score 0.69). These findings show the feasibility of recognizing key football actions and head scans with wearables and highlight promising sensor locations.
Dette studie kombinerer inertimåleenheder (IMU'er) – små kropsbårne bevægelsessensorer – med principper fra aktivitetsgenkendelse (Human Activity Recognition) for automatisk at genkende almindelige fodboldhandlinger og pege på, hvor sensorer bør placeres for at give mest information. Arbejdet blev delt i to. Først blev fodboldhandlinger som pasning, dribling, førsteberøring og positionering forudsagt med et to-vejs Long Short-Term Memory neuralt netværk (LSTM), som er velegnet til tidsserier. Derefter blev head scans (hovedskanninger, der kan forekomme samtidig med andre handlinger) forudsagt separat med et dybt neuralt netværk (DNN). I alt deltog 17 spillere (14 mænd og 3 kvinder). Data blev først delt 50/50 i mærkede og umærkede datasæt; de mærkede data blev derefter delt 80/20 i trænings- og testdata. Alle data blev normaliseret og klassebalanceret med ADASYN. Der blev udregnet 5.250 statistiske tidsdomænefeatures i et glidende vindue på 200 ms med 50 % overlap og efterfølgende reduceret med hovedkomponentanalyse (PCA), så >95 % af variansen blev bevaret. En semiovervåget, usikkerhedsbevidst pseudo-labeling-metode blev brugt for at reducere tidsforbruget til manuel mærkning. LSTM-modellen gav pæne resultater for fodboldhandlinger med en krydsvalideringsscore på 0,74 og en F1-score på 0,74. DNN'en til head scans klarede sig en smule bedre, primært fordi der var færre klasser, med en krydsvalideringsscore på 0,79 og en F1-score på 0,78. Placeringen, der bidrog mest til LSTM'ens præstation, var sensoren på højre læg (F1-score 0,65). For DNN'en var sensoren på hovedet vigtigst (F1-score 0,69). Resultaterne viser, at det er muligt at genkende centrale fodboldhandlinger og head scans med bærbare sensorer og peger på, hvilke sensorplaceringer der er mest nyttige.
[This apstract has been rewritten with the help of AI based on the project's original abstract]