Invisible Map

Overview

In summer 2021, I worked in the Olin College Crowdsourcing and Machine-Learning (OCCaM) Lab under advisor Paul Ruvolo. OCCaM Lab co-designs apps with and for the blind and visually impaired (B/VI) community. Their major app, Clew, uses Apple’s ARKit to guide users back along a path they’ve walked with a sighted guide.

My project focused on Invisible Map, a prototype navigation app extending Clew to support full indoor navigation without a sighted guide. The idea: building owners can create a single map of their building which could then be shared with B/VI users for independent navigation.

By the end of the summer, we delivered a reliable proof-of-concept demo, improving SLAM accuracy and showing Invisible Map’s potential to expand autonomy for blind and visually impaired users.

Challenges

Invisible Map’s core challenge was a Simultaneous Localization and Mapping (SLAM) problem: when someone was making a map, the app needed to track their location in the building while simultaneously constructing the map of the building.

• Drift - compounded sensor error degraded accuracy over time.
• Outliers in AprilTag detection - misread tags skewed maps, sometimes by meters.

My Contributions

1. Weighting Error Sources

• Investigated G2O optimization for SLAM with 12 error weights (x, y, z, yaw, pitch, roll for two detection types).
• Implemented a genetic algorithm to tune weights, but lacked ground truth for validation.
• Discovered that inconsistencies came from different magnitudes in weight vectors. Normalizing standard weights resolved ranking issues and improved reliability.

2. Improving AprilTag Detection

• Camera-based AprilTag detection introduced severe outliers in orientation, especially at the oblique angles common with this use case.
• Integrated LiDAR raycasting to flatten AprilTag readings onto walls, leveraging LiDAR’s accurate surface orientation.
• This dramatically reduced rotational errors and produced much cleaner maps.

Results

By the end of the summer, my project partner and I transformed Invisible Map from a struggling prototype into a working proof-of-concept demo. With improved weighting strategies and LiDAR-enhanced AprilTag detection, maps were significantly more accurate and robust.

Future Work

We also began tackling path optimization. The existing "breadcrumb" approach required users to exactly retrace the map creator’s path, leading to inefficient routes. Our contribution was a LiDAR-based algorithm to detect whether two points were connected by a floor (with no walls between them).

This raycasting method showed promise for detecting overlapping regions and enabling shortcuts, though it remained as a demo for future researchers to integrate into Invisible Map.