Synopsis: Experimental evaluation of unsupervised indoor localization.
Level: Summer internship or BS/MS project
Due to their aptitude in depicting complex dependencies, neural networks are a promising candidate for indoor localization. Omnipresent phenomena such as multi-path signal propagation, shadowing and device noise introduce high noise in distances inferred from radio frequency signals, and make conventional geometry-based methods fail even in simple environments.
This project builds on a successful previous proof-of-concept semester project. We have developed an unsupervised indoor localization method which works on distance estimates obtained from WiFi and Bluetooth beacons, and relative movement measurements obtained from inertial measurement units (IMUs). The developed network works well in simulation, trained on ca. 200 two-minute long paths without labels (=grund truth positions).
This amount of training data being unpractical, the goal of this semester project is to test the ability of the network to learn from only a few (<5) paths, by maximally reducing the network complexity and using a more sophisticated training scheme. Then, the network can be tested on a number of publicly available indoor localization datasets, as well as on our own setup acquiring WiFi RTT, Bluetooth RSSI and IMU measurements. The project is concluded by concisely evaluating the obtained accuracy and comparing with state-of-the-art indoor localization algorithms.
Deliverables: High-quality code and a short project report including experimental results.
Prerequisites: Solid machine learning basics, good programming skills (preferably experience with PyRorch). Curiosity and willingness to learn.
Type of Work: 20% theory, 40% coding, 40% experimental validation.