Tegevuse tuvastamine kiirendusandurite abil
Date
2019
Authors
Journal Title
Journal ISSN
Volume Title
Publisher
Abstract
Automaatsel tegevuse tuvastamisel on palju rakendusi, iseäranis tervise valdkonnas. Erinevate igapäevaeluliste tegevuste mõõtmine on kasulik, sest see võimaldab saada teavet terviseseisundite kohta, nagu näiteks ülekaalulisus, insult või kukkumine. Veelgi enam, erinevate kasutajasõbralike kantavate seadmete laialdane levik võimaldab koguda kolmeteljelise kiirendusanduri andmeid mittesegavalt ja diskreetselt. Käesolevas töös on tegevuse tuvastamiseks kasutatud kiirendusanduri andmed on pärit projektist SPHERE [1]. Kiirendusmõõtmised on tehtud nelja kantava seadmega, mis olid kinnitatud katseisiku randmetele ning jalgadele.Töö võrdleb otsustusmetsa (random forest) ning pika lühiajalise mäluga (LSTM) tehisnärvivõrkude võimet tuvastada 9 siseruumi tegevust: hambapesu, söömine, hambaniiditamine, riietumine/lahtiriietumine, (toidu) segamine, (toidu) pealemäärimine, kõndimine, käte pesemine, kirjutamine. Lisaks laiendatakse tegevuste hulka teadmata tegevusega. Suuremat tähelepanu pööratakse järgmistele teemadele: tunnuste eraldamine, kiirendusanduri aegrea tükeldamine, parameetrite ja hüperparameetrite häälestamine, mudeli treenimine, mudeli hindamine ning üldistusvõime. Tulemused näitavad, et kiirendusanduril põhinevaid ekstraheeritud tunnuseid kasutav otsustusmets ületab tuvastusvõimelt pisut kiirendusanduri mõõtetulemusi muutmata kujul kasutavat LSTM-võrku, ning seda nii 9 tegevuse tuvastamisel kui ka peale teadmata tegevuse lisamist.
Activity recognition is considered to have a wide range of applications, especially in the health sector. The assessment of different activities of daily living is useful because it can highlight information related to a specific health condition such as obesity, overweight, stroke, or fall. Moreover, the prevalence of different user-friendly wearable devices enables collecting tri-axial accelerometer data in a non-intrusive and discrete manner.The accelerometer data used for activity recognition in this thesis is provided by SPHERE [1]. The accelerometer readings are recorded from four wearables attached on a single person's hands and legs.This thesis compares the capabilities for activity recognition of the random forest model and the long short-term memory neural network to discern among 9 in-door activities including brushing teeth, eating a meal, flossing, getting dressed/undressed, mixing (food), spreading (food), walking, washing hands, writing. In addition, the list of activities is extended with an unknown activity. Greater focus is given on the following topics: feature extraction, segmentation of the time-series accelerometer data, parameter and hyper-parameter tuning, model training, model evaluation and generalization capability. The results suggest that the random forest model using the accelerometer-based extracted features slightly outperforms the long short-term memory neural network using raw accelerometer data when the activity recognition task is limited on the 9 chosen activities, and, additionally, when the unknown activity is included.
Activity recognition is considered to have a wide range of applications, especially in the health sector. The assessment of different activities of daily living is useful because it can highlight information related to a specific health condition such as obesity, overweight, stroke, or fall. Moreover, the prevalence of different user-friendly wearable devices enables collecting tri-axial accelerometer data in a non-intrusive and discrete manner.The accelerometer data used for activity recognition in this thesis is provided by SPHERE [1]. The accelerometer readings are recorded from four wearables attached on a single person's hands and legs.This thesis compares the capabilities for activity recognition of the random forest model and the long short-term memory neural network to discern among 9 in-door activities including brushing teeth, eating a meal, flossing, getting dressed/undressed, mixing (food), spreading (food), walking, washing hands, writing. In addition, the list of activities is extended with an unknown activity. Greater focus is given on the following topics: feature extraction, segmentation of the time-series accelerometer data, parameter and hyper-parameter tuning, model training, model evaluation and generalization capability. The results suggest that the random forest model using the accelerometer-based extracted features slightly outperforms the long short-term memory neural network using raw accelerometer data when the activity recognition task is limited on the 9 chosen activities, and, additionally, when the unknown activity is included.