20.500.12556/DKUM-72257 Razpoznavanje vrste aktivnosti uporabnika in prevoznega sredstva s pomočjo vgrajenih senzorjev mobilne naprave User activity and transportation mode recognition using embedded mobile device sensors Cilj tega zaključnega dela je izboljšati detekcijo vrst aktivnosti uporabnika in prevoznih sredstev v mobilni aplikaciji MobilitApp. Z uporabo znanih tehnik obdelave signalov in razpoznavanja vzorcev smo razvili dve metodi za detekcijo aktivnosti uporabnika. Prepoznavali smo naslednje aktivnosti: mirovanje, hoja, tek, vožnja s kolesom, motorjem, avtom, avtobusom, metrojem, vlakom in tramvajem. Zgolj z uporabo vgrajenih senzorjev mobilne naprave smo v prvem pristopu zgradili vektor značilnic, ki smo ga uporabili za učenje odločitvenih dreves. V drugem pristopu smo na vhod nevronskih mrež poslali neobdelane podatke iz pospeškometra. Zajemanje podatkov smo zabeležili z video kamero z namenom natančnejše izločitve posameznih delov aktivnosti iz signala. Izvedli smo še študijo, če je kontrolirano zajemanje podatkov z video kamero dejansko potrebno. Naše pristope smo validirali na manjši lastni podatkovni zbirki, ki obsega okrog sedem ur aktivnosti. Podatke smo zajeli s pomočjo treh prostovoljcev in štirih različnih mobilnih naprav. Rezultate smo ovrednotili z različnimi metrikami, s poudarkom na metriki »F1 score«. S pristopom Random Forest je bila metrika »F1 score« 94 %, kar je bolje kot uspešnost obstoječih rešitev iz literature. Naš pristop na osnovi nevronskih mrež se je izkazal nekoliko slabše (»F1 score« je bil 91 %), a še vedno bolje kot najsodobnejše metode iz literature. This study aims to improve the recognition of user activity type and transportation mode in the mobile application MobilitApp. By using known signal processing and pattern recognition techniques, we developed two methods to recognize user activities. We were recognizing the following activities: stationary, walking, running, driving on a bicycle, a motorcycle, a car, taking the bus, the metro, the train and the tram. By using embedded mobile device sensors only in the first approach, we built a feature vector that we used to train decision trees. In the second approach, we provided raw data from the accelerometer to the input of a neural network. Data capture was recorded with a video camera in order to precisely identify the individual parts of the activity in the signal. We also did a study to ascertain whether a controlled capture of data with a video camera was, in fact, necessary. We validated our approaches on a smaller database of our own, which covers about seven hours of activity. The data was captured with the help of three volunteers and four different mobile devices. The results were evaluated with different metrics, with an emphasis on the 'F1 score' metric. With the Random Forest approach, the 'F1 score' metric was 94%, which is better than the success of existing 'state-of-the-art' solutions. The performance of our approach based on neural networks was slightly lower ('F1 score' was 91%), but still better than most of the recent methods in the literature. obdelava signalov razpoznavanje vzorcev aktivnost uporabnika prevozno sredstvo signal processing pattern recognition user acitivity transporation mode true true false R. Slamek Slovenski jezik Angleški jezik Magistrsko delo/naloga 2018-09-17 15:42:59 2018-11-08 11:31:55 2022-08-01 22:11:44 2019-07-11 19:12:24 2018 0 [Maribor 0 0000-00-00 NiDoloceno NiDoloceno NiDoloceno 0000-00-00 0000-00-00 0000-00-00 004.93(043.2) 21873686 URN:SI:UM:DK:R7SJXHEU MAG_Slamek_Rok_2018.pdf MAG_Slamek_Rok_2018.pdf 1 1B9F336569188A67D055BB8F462A7D32 1c060556db58685672855447e6738cb87a4519a4092dbea2e3ba88a4f13289ce 5764d52f-7c10-11eb-bb7a-00155d0001ca 20.500.12556/dkum/ee481f3f-150f-4dee-91b7-851ab5830376 https://dk.um.si/Dokument.php?lang=slv&id=129660 Fakulteta za elektrotehniko, računalništvo in informatiko 0 0 0