Physik | Technik


Mark Marolf, 2002 | Baden, AG


Indoor positioning systems (IPS) can automatically track the movements of indoor objects in real time. The goal of this project was to develop the hardware and software necessary to create a cheap, high-performance IPS from scratch. A system was constructed that automatically approximated the position of an object using Bluetooth Low Energy (BLE) signals. Using the log-distance path loss model, the received BLE signal strengths were used as a proxy for the distance separating two devices. Distance estimates were used to trilaterate positions. Research and development were done on theoretical and practical improvements. The resulting IPS served as a viable alternative for GPS, which generally lacks coverage indoors.


Is it possible to create the hardware and software for an IPS? Additionally, how can its accuracy and feasibility be improved in theory and practice?


First, I developed software to efficiently collect, store and evaluate BLE signal strengths from cheap microcontrollers. After collecting data at fixed intervals between devices, the software could calculate room-specific signal path loss coefficients using linear regression. These were fed into the log-distance path loss model, which mapped signal strengths to distances. From these distance values, position estimates were computed using trilateration. In CAD, I designed 3D printable cases and stands for the microcontrollers to achieve a specific spatial placement. This reduced noise in the BLE signals. On the software side, I improved the reliability of distance and position estimates algorithmically. For this, I conducted over 70 experiments attempting to enhance accuracy and performance. To demonstrate a concrete use case of an IPS, I finally set up the IPS as a backend for a web application. It allowed for a shop owner to view customer flows and monitor how long they looked at specific products.


The IPS had an average root mean squared error (RMSE) of 81 cm in a 4.17 m by 5.40 m room without obstructions. When data from rooms of various sizes with differing path loss properties were included, the overall average position estimate RMSE was 94 cm. Generally speaking, this was sufficient to estimate in which area of a room a device was. Positioning latency was approximately one second, with updates made at a frequency of 10 Hz. The custom cases and stands reduced signal strength measurement standard deviations by 53%. Doubling the count of BLE microcontrollers in the room from 4 to 8 decreased the average position estimate RMSE by 38%. Path loss coefficients remained valid a month after calibration as indicated by statistically insignificant changes to RMSE averages.


The accuracy and precision of the results were comparable to literature. In this regard, I was relatively satisfied with the capabilities of the system. The lackluster accuracy can almost certainly be attributed to the high variance in the signal strength measurements. Basing calculations on data of limited quality put constraints on the performance of the IPS. More elaborate processing techniques would probably not have yielded vastly improved accuracy. Intrinsic shortcomings in the measurement data beg the question of how accurate systems of this type even can be. However, as shown in the store web application, certain use cases don’t necessarily require higher accuracy. More tangible improvements could have been attained by improving the practicality of the coefficient calibration sequence and shrinking the size of the antennae.


In conclusion, the initial question whether it is possible to create the hardware and software for and IPS was affirmed. In the future, it would be interesting to try to automate the calibration sequence and try other approaches to indoor positioning. On a personal level, the many setbacks and problems allowed me to learn a lot and gather valuable experience. My main takeaways are the following: Firstly, a large amount of time should be invested in research. Secondly, premature optimization is to be avoided at all costs. Thirdly, it is important to validate the correctness of formulas by running mock data through them. Finally, to get a more complete evaluation of results, central tendency measures should be considered in addition to dispersion measures. These insights will certainly be valuable for future projects.



Würdigung durch den Experten

Prof. Zeno Stössel

Mit grossem Einsatz und Begeisterung wollte Mark Marolf das seit Jahren bestehende Problem einer zuverlässigen und kostengünstigen Indoor Localization mit Bluetooth Low Energy Beacons lösen. Dabei wurde er mit der sehr unzuverlässigen Distanzbestimmung basierend auf der Signalstärke der Beacons konfrontiert. Mit guten Ideen versuchte er die Genauigkeit zu verbessern. Alle Schritte wurden mit vielen Messungen hinterlegt. Die Dokumentation ist gut lesbar und nachvollziehbar. Mark bearbeitete die auftretenden Probleme eigenständig, mit viel Ausdauer und sehr originell.


sehr gut

Sonderpreis ETH Student Project House




Aargauische Kantonsschule Baden
Lehrer: Daniel Süsstrunk