Inspiration

We’ve heard of countless complaints about navigating in the maze-like Mosse humanities building, and in general, students spend a lot of time finding rooms inside unfamiliar buildings, whether students are going to an exam, office hours, etc. In the navigation industry, there are a lot of options to navigate outside buildings, but there are no reliable and universal tools to navigate indoors. Existing indoor navigation tools are often developed for a single building or location, and the tools require inaccessible data to work, so we set off to develop a tool to make indoor navigation possible for all enclosed spaces that allows users to contribute to the mapping of an indoor space.

What it does

MadMaps allows users to navigate to different rooms inside buildings, whether it’s the restroom, classrooms, offices, etc. MadMaps uses vision-based localization and floor-plan aware routing to calculate routes for the user, so the functionality is nearly identical to Google Maps and other navigator tools, but for indoors.

MadMaps uses a graph data structure to map out a building. Edges and weights are determined by correlating the pixel lengths of the floor plan with real life distances.

To determine a user's position within a building, MadMaps uses ARCore Cloud Anchors. These anchors are created by scanning visually distinctive areas with a smartphone, allowing ARCore to capture recognizable features of the surrounding environment. Each anchor is associated with a known location and orientation on the building's floor plan.

When another user approaches the area, ARCore can recognize the previously scanned environment and resolve the anchor. The app then calculates the user's position relative to that anchor and translates it into coordinates on the floor plan. As the user moves, ARCore continuously tracks their position, enabling real-time indoor navigation.

We enforce data safety in our backend. Some examples of the rules that we enforce is that edges cannot pass through walls, nodes must be within the walls of the building, and there can be no negative edges.

After MadMaps builds the undirected graph mapping of the building, we can use shortest-path algorithms for routing. We decided to use the A* (A star) algorithm since it is more efficient than Dijkstra’s. We used a heuristic that we call Landmark ALT (A star, Landmark, Triangle Inequality) that uses the triangle inequality to calculate a lower bound (making sure that our heuristic isn’t larger than the actual distance from the destination). Geometric distance (like Euclidean distance heuristic) performs poorly when there are walls, detours, stairs, and unreliable coordinates, hence we decided to use Landmark ALT.

MadMaps also allows users to upload a floor map and extract essential data for it for MadMaps to build an initial map. We integrated Databricks Model Serving into our floor-plan extraction workflow, using Llama 4 Maverick to generate map annotations for review.

How we built it

We built MadMaps Flutter, ServerPod, and DataBricks. Flutter and Serverpod integrate seamlessly, providing a convenient frontend and backend system. We use DataBricks for our AI layer, where it can read a floor plan and generate an initial graph-based representation of the floor plan. The graph-based representation will change as nodes are added by users.

Challenges that we ran into

It was difficult to extract data from floor maps properly, and there were a lot of little details to be ironed out in our localization. Another big issue was handling unsafe data, so we had edges that went through walls.

Accomplishments that we’re proud of

We’re proud of developing a working version of MadMaps that, in the future, can be extended to work with any building.

What we learned

Building backend systems with serverpod, localization techniques, and effective teamwork strategies.

What’s next for MadMaps

Effectively fine-tune a vision language model to allow users to upload a floor plan and extract a map of it. We've tried, but ran out of time and didn't get a good result. Expand security features for crowdsourced points and build some sort of administrative layer to prevent any troll data points

Built With

Share this project:

Updates

Submission history