Merjenje razdalje na osnovi eholokacije z uporabo mobilnega telefona

IDZIG, JURE (author), Stančin, Sara (mentor)

Cilj, zastavljen v diplomski nalogi, je ugotoviti omejitve in določiti natančnost merjenja razdalje preko odboja zvočnega signala od stene. Signal je bil ustvarjen in zajet z zvočnikom in mikrofonom, vgrajenima v Android telefon. Natančneje je preko klasifikacijskega modela nevronske mreže bila ugotavljana uspešnost prepoznave razdalje pri meritvah z milimetrskimi, centimetrskimi in desetcentimetrskimi razmaki. Namen je bil ugotoviti maksimalno razdaljo, s katere je mogoče še uspešno klasificirati razdalje preko merjenja odboja, minimalno razdaljo, pri kateri nevronska mreža še uspešno ločuje odboje med seboj, in primerjati uspešnost metode klasifikacije z uspešnostjo metodo merjenja razdalje preko vrhov korelacijske funkcije. V sklopu diplomske naloge je bila v programu Android Studio najprej izdelana aplikacija, ki proizvaja željene signale in zajema posnetke odboja. Posnetki so bili zajeti s 44.100 Hz vzorčne frekvence in shranjeni na telefon srednjega cenovnega ranga v .pcm obliki, da bi ohranili maksimalno natančnost meritev. V programu Matlab je bila sprogramirana in preizkušena nevronska mreža, ki je osnovana na kodi iz spletnega tečaja strojnega učenja. V zadnjem koraku so zajeti podatki bili korelirani z oddanim signalom in nevronsko klasifikacijski model je bil naučen prepoznave razdalj. Pogoji zraka med zajemi posnetkov nismo bili spremljani. Rezultati naučenega modela pri meritvah z določenih razdalj kažejo precejšno uspešnost. Z zadostno količino podatkov ter raznolikostjo pogojev v prostoru med zajemanjem le teh bi model potencialno lahko z osemmilimetrsko natančnostjo napovedoval razdalje med 9 in 160 centimetrov oddaljenosti od stene. Pri večjih razdaljah je natančnost manjša – v tem primeru se uporabi metodo vrhov korelacijske funkcije, ki omogoča do desetcentimetrsko natančnost napovedovanja. Področje raziskave omogoča še veliko prostora za nadgradnje ter opravljanje dodatnih meritev in testov, saj je v zraku še veliko spremenljivk, ki bi lahko vplivale na rezultate le te.

• Language: Slovenian
• Keywords: merjenje razdalj, zvok, odboj, nevronska mreža, klasifikacija, Android aplikacija, Matlab
• Work type: Undergraduate thesis
• Organization: FE - Faculty of Electrical Engineering
• Year: 2020
• PID: 20.500.12556/RUL-120269
• Publication date in RUL: 17.09.2020

Secondary language

• Language: English
• Title: Echolocation-based distance measurements using mobile phones

Abstract

The objection of this senior thesis is to find out the limits and set the accuracy of measuring distance, using the method of echolocation with the sound signal being emitted and received by a smartphone. The sound signal was created and received using the speaker and microphone, both inbuilt into the Android smartphone. Classification neural network was determining the accuracy of distance recognition in measurements with millimeter, centimeter, and ten-centimeter intervals between each. The purpose was to determine the maximal distance, at which it is still possible to successfully classify distance using the method of echolocation via smartphone, the minimal distance, at which the neural classification network still differentiates the sound rebounds amongst themselves, and to compare accuracy of both classification method and the method of measuring distance between peaks of correlation function. Firstly, an application was created in Android Studio, whose function is to emit desired sound signals and receive the sound signal rebounds. The recordings were captured using the 44.100 Hz sampling frequency and then saved onto a middle price range smartphone in a .pcm file format in order to preserve a maximal accuracy of the measurements. Furthermore, a neural classification network was created in Matlab on a basis of a pre-existing code. In the last step of the research all the recorded data was correlated with the emitted sound signal and the neural classification network was taught to recognize the different distances. Air conditions in the research environment were not measured and recorded. The results of the research show a considerable among of accuracy. With an adequate amount of data and the conditions in the environment at the time of measurement-taking being diverse enough, the neural classification model could potentially successfully predict distances between 9 and 160 centimeters with an eight-millimeter accuracy. When trying to determine distances pass 160 centimeters the predictions are less accurate – in that case the method of measuring distance between peaks of correlation function is used, which enables a ten-centimeter accuracy of prediction. The field of this research still enables a lot of space for improvements, more additional measurements and testing, since there are different air conditions which have to be taken into account and can interfere with the results.

• Keywords: distance measuring, sound, rebound, neural network, classification, Android application, Matlab