With the development of the Internet of Things, objects with computing capabilities are becoming part of the environment. Real-time context recognition is key to seamlessly integrating ubiquitous intelligence into our daily lives, as it allows smart objects to adapt their behavior to the situation. In this thesis, we study the feasibility for mobile phones to automatically recognize the elevator context from on-board sensor data without external infrastructure. By realizing such a real-time trigger, a wide range of opportunities may be enabled for next-generation elevator systems, such as just-in-time information display or crowdsensed installation monitoring for predictive maintenance. To automatically label multi-modal sensor data a posteriori, we apply a function fitting algorithm to exploit the characteristic signature of vertical acceleration within the elevator car. From this context segment, we heuristically obtain labels for the contexts of in-car/pre-acceleration and the null-class of outside the elevator context entirely. Following a window-based supervised learning approach, we compare the performance of support vector machine, random forest, gradient tree boosting, convolutional and recurrent neural networks. The algorithms are trained on a dataset including 22 hours of sensor data collected by 15 people recording trips on approximately 15 different elevator installations. The XGBoost implementation of gradient tree boosting outperforms all the other algorithms, achieving 0.92 in area under ROC curve score and detecting elevator entries and trips with an average delay of 0.7 seconds and 1.7 seconds, respectively. Interestingly, it also outperforms deep learning, we believe because of the relatively small size of the dataset. Finally, we empirically validate our real-time recognition pipeline on an Android device, proving that modern smartphones are capable of recognizing the elevator context through their sensors.
Con lo sviluppo dell’Internet of Things oggetti capaci di eseguire calcoli sono diventati parte dell’ambiente. Il riconoscimento del contesto in tempo reale è fondamentale la loro integrazione nelle nostre vite quotidiane, in quanto permette agli oggetti di adattare il loro comportamento alla situazione. In questa tesi studiamo la fattibilità per i telefoni cellulari di riconoscere il contesto-ascensore attraverso i dati raccolti dai propri sensori, senza il bisogno di infrastrutture esterne. La realizzazione di questo meccanismo di attivazione in tempo reale aprirebbe a numerose opportunità per la prossima generazione di ascensori. Per esempio, consentirebbe la visualizzazione di informazioni prima dell’inizio del viaggio o il controllo dello stato di funzionamento della macchina. Per etichettare automaticamente i dati dei vari sensori, una volta raccolti, utilizziamo un algoritmo di function fitting che sfrutta la traccia caratteristica dell’accelerazione verticale percepita all’interno dell’ascensore. A partire da questa segmentazione otteniamo etichette per i contesti "in ascensore/prima dell’accelerazione" e per la classe nulla "fuori dall’ascensore". Seguendo un approcio supervised, basato sulla suddivisione del segnale in finestre temporali, confrontiamo le performance di support vector machine, random forest, gradient tree boosting, convolutional e recurrent neural networks. Il training degli algoritmi viene effettuato su un dataset di 22 ore di dati raccolti da 15 persone, comprendene viaggi su 15 diversi ascensori. XGBoost, una libreria che implementa gradient tree boosting, supera tutti gli altri algoritmi testati, raggiungendo un valore di area under ROC curve pari a 0.92 e individuando l’ingresso in ascensore e l’inizio del viaggio con un ritardo medio di rispettivamente 0.7 e 1.7 secondi. Supera in accuratezza anche gli algoritmi di deep learning, ipotizziamo per via delle ridotte dimensioni del dataset. Infine, validiamo empiricamente l’architettura per il riconoscimento del contesto-ascensore in tempo reale su un device Android.
Real-time recognition of industry-specific context from mobile phone sensor data
CHIAPPA, ALBERTO
2017/2018
Abstract
With the development of the Internet of Things, objects with computing capabilities are becoming part of the environment. Real-time context recognition is key to seamlessly integrating ubiquitous intelligence into our daily lives, as it allows smart objects to adapt their behavior to the situation. In this thesis, we study the feasibility for mobile phones to automatically recognize the elevator context from on-board sensor data without external infrastructure. By realizing such a real-time trigger, a wide range of opportunities may be enabled for next-generation elevator systems, such as just-in-time information display or crowdsensed installation monitoring for predictive maintenance. To automatically label multi-modal sensor data a posteriori, we apply a function fitting algorithm to exploit the characteristic signature of vertical acceleration within the elevator car. From this context segment, we heuristically obtain labels for the contexts of in-car/pre-acceleration and the null-class of outside the elevator context entirely. Following a window-based supervised learning approach, we compare the performance of support vector machine, random forest, gradient tree boosting, convolutional and recurrent neural networks. The algorithms are trained on a dataset including 22 hours of sensor data collected by 15 people recording trips on approximately 15 different elevator installations. The XGBoost implementation of gradient tree boosting outperforms all the other algorithms, achieving 0.92 in area under ROC curve score and detecting elevator entries and trips with an average delay of 0.7 seconds and 1.7 seconds, respectively. Interestingly, it also outperforms deep learning, we believe because of the relatively small size of the dataset. Finally, we empirically validate our real-time recognition pipeline on an Android device, proving that modern smartphones are capable of recognizing the elevator context through their sensors.| File | Dimensione | Formato | |
|---|---|---|---|
|
2018_07_Chiappa.pdf
non accessibile
Descrizione: Thesis text
Dimensione
4.18 MB
Formato
Adobe PDF
|
4.18 MB | Adobe PDF | Visualizza/Apri |
I documenti in POLITesi sono protetti da copyright e tutti i diritti sono riservati, salvo diversa indicazione.
https://hdl.handle.net/10589/141690