WhereAmI? A Mixed Reality application for Indoor Localization

Team: Árni Bjarnsteinsson, Fabian Blatter, Matěj Mrázek, Ross Roessler

Download the Final Report See the Code on Github

Abstract

This was a group project for the Mixed Reality class at ETH. We built an Augmented Reality for a Magic Leap 2 device that allowed the user to locate their position indoors and navigate in a previously mapped environment. We used the Hierarchical-Localization algorithm on reference data inside one of our campus buildings at ETH from the LaMAR dataset.

My Contributions

• Cleaning and preprocessing the data we needed from the LaMAR dataset.
• Updating the Hierarchical-Localization algorithm to run and make predictions on that data.
• Experimenting with different algorithms and parameters to get the best localization results.

Analysis

Challenges

• Terabytes of data requiring different transformations - we weren’t even sure at first which data we could use
• Relatively sparse ground truth global trajectory data
• Poor predictions from COLMAP

Learnings

• Sending fewer images to COLMAP actually worked better
• Integrating different ground truth data sources helped significantly
• Visualization tools for debugging were extremely helpful

My teammate Árni built the visualization on the bottom right, which showed the ground truth locations and the prediction location. It was super helpful in debugging.

Real world on the left. What the user sees is on the right, and the white line is the navigation path they're following.