Overview
For my master’s thesis at TU Berlin, I developed a complete pipeline that transforms robot-scanned 3D data into a functional neural network for indoor localization. The system eliminates the need for manual data collection by training exclusively on synthetic images, achieving practical accuracy on a real-world office building dataset.
Background
Indoor localization is fundamental to many applications, from assistive navigation for people with visual impairments to augmented reality experiences. While GPS works great outdoors, it fails indoors due to signal reflections and building materials blocking satellite signals. Traditional approaches like WiFi fingerprinting or visual markers work, but they either lack accuracy or require extensive infrastructure installation.
Neural networks offer a promising new path, using just a device camera to determine precise location and orientation.
The challenge here is data collection, as networks need thousands of pose-annotated training images to function properly. Collecting this data manually is time-consuming and prone to gaps that may not be obvious until after the network is put into use.
Another open question is how well localization networks can perform in large, self-similar environments, like the corridors of an office building. Most existing research focuses on smaller, more distinctive spaces, leaving real-world scalability unproven.
My thesis addresses both challenges: building an end-to-end pipeline that generates synthetic training data from existing robot scans (specifically LIDAR point clouds and 360° images captured during routine mapping), while also validating whether this approach actually works at scale in a large office environment.