Project · Maker
AccessMate: a camera that tells you what is around you
AccessMate started as an app that uses a camera and computer vision to recognize objects and describe them to people who are blind or have low vision. It is now turning into a helmet that watches in every direction and warns the wearer about what is coming. The helmet design holds a Taiwan utility model; I am still designing the full physical build.
- Role
- Developer & Inventor
- When
- Oct 2024
- Status
- App built; helmet hardware being designed
The problem
A white cane finds what is on the ground a step ahead. It says nothing about a bicycle approaching from the side, a sign at head height, or which door in a row of doors is the pharmacy. I wanted a tool that could look around for someone and say, in plain words, what it sees.
Who it's for
The first users I had in mind are people with visual impairments moving through busy streets. The helmet idea came later, in Taiwan, where I watched motorcyclists weave through traffic with cars and scooters closing in from every side. A rider can only look one way at a time. The same helmet that describes a street to a blind pedestrian could warn a rider about the car in their blind spot.
What I built
The first version is an app. It points the phone or computer camera at the scene, runs a real-time object detector, and reads out what it finds. That app is what earned recognition in the 2024 Congressional App Challenge.
From there I designed a wearable version: a helmet with cameras covering 360 degrees and the processing built in, so it works without a phone in hand. The design, "Smart AI helmet real-time environmental perception & description system," was granted Taiwan Utility Model M676830 in November 2025. (A utility model is Taiwan's patent for a practical device design.)
I am still designing the physical build, which adds radar and on-board AI acceleration, so I don't have photos of a finished helmet yet.
What happened
- Nov 2025 Taiwan Utility Model M676830 granted for the smart helmet design
- Nov 2024 Congressional App Challenge recognition for the AccessMate app
The work continues in a research setting. Since September 2026 I have been doing mentored research on user-centric AI for assisting the visually impaired with Lilian Chen at Stanford's Center for Artificial Intelligence in Medicine & Imaging (AIMI). See Research.
Under the hood
The app runs a pretrained object detector on live camera frames and converts the labels into speech. In August 2025 I packaged a real-time detector built on Ultralytics YOLOv11 (Python, PyTorch, OpenCV) so it runs fully offline, with the model weights and dependencies bundled for a machine that has no internet. Offline matters for a helmet: it can't depend on a cell signal in a tunnel or on a mountain road. This uses off-the-shelf weights; I have not trained a custom model or measured accuracy on my own data yet.
The planned helmet hardware:
- Raspberry Pi 5 as the main computer
- An AI accelerator board so detection runs fast enough on a battery-powered device
- A camera for object recognition
- A 24 GHz radar module to measure how fast something is approaching and estimate time to collision, which a camera alone does poorly in rain or at night
The open questions are the ones that decide whether anyone would wear it: weight, battery life, heat, and how to give warnings through sound without burying the wearer in chatter.