M.Sc. Topic / Guided Research: Sensor Data Analysis Inferring Group Proximity at Different Scales

 

We live in a more and more connected world, sharing information with people around the world. Our goal is to enrich the way how the user groups are formed which are sharing information. In our case, we are interested in a proximity-based grouping of nearby users based on environment-specific sensor data [1]. For instance, for an easy sharing of information with people using the same subway or keeping tourist groups together and inform them about points of interest. Therefore, we have a working prototype to detect other nearby devices in the same environment. We use Wi-Fi signals [2], ambient audio similarity and ultrasound in periods of silence. As a result, we are able to autonomously group devices based on their perceived environment. The goal of the master thesis is to gain further insights about which sensor data is useful for proximity-based device grouping. Therefore, we have collected sensor data including information about accelerometer (detect transportation mode), barometer, Bluetooth encounters, GSM, location data (GPS, Network), magnetometer, and Wi-Fi networks from 126 devices in a time period of several months.

Possible aspects of the data analysis

  • Time dependency of sensor data
    • At which time does the proximity prediction works better
  • Energy consumption of sensors
    • Which subsets of sensor data produce which accuracy vs. energy
  • Device diversity: different sensors in different smartphones produce different results
    • Learn a transformation over multiple different sensors to enhance the robustness of the machine learning prediction

In general, the following questions are to answer:

  • Which modalities are most meaningful for proximity detection in local environments?
  • Which sensor data provides which distance granularity (e.g., Bluetooth encounters in the range of 20m and Wi-Fi signals in the range of 100m)? 

Requirements

  • Experience with Python or R
  • Familiarity with signal analysis, machine learning 

Contact

If you are interested into the thesis, please send me your CV and exams transcript.

Michael Haus – haus@in.tum.de

References

[1] Predicting Location Semantics Combining Active and Passive Sensing with Environment-Independent Classifier

[2] Inferring Person-to-person Proximity Using WiFi Signals