Embedded · Signal processing · Applied AI
FSP: Foot Pressure Mat
A low-cost foot-health screening product: sensor mat, ESP32 firmware, pressure heatmaps, guarded AI analysis and a PDF report.
- Status
- Working prototype
- Where
- Innovizia Technologies
- My role
- Owned end-to-end technical development
- 300
- people in the real test dataset
- Concept → PCB
- from first idea to a manufactured, assembled board
- Low-cost
- designed to be cheap enough to reach a wide audience in India
Overview
Foot assessment is usually done with expensive equipment. FSP's goal was a screening product cheap enough to manufacture at scale in India. I owned the technical side from the first concept to a working prototype: the sensing circuit, the mat itself, firmware, signal processing, the analysis pipeline and the report.
- 01 · Sense Resistive matrix Flexible sensor mat, layered and assembled by hand
- 02 · Scan ESP32 + multiplexers Voltage dividers, Wi-Fi and USB-serial, battery powered
- 03 · Clean Normalise + de-ghost Median-filter-based algorithm removes matrix ghosting
- 04 · Map Foot regions Elliptical regions on a foot silhouette → heatmap
- 05 · Decide Rules engine Condition table maps findings to likely conditions
- 06 · Explain LLM + foot images Personalised text; rules stay in charge
- 07 · Report PDF Results, risks, exercises, insole advice
How a sensor matrix gets cleaned up
This is an illustration with simulated data. It shows the ideas, not real FSP readings or the production algorithm.
What I did
- Designed the sensing circuit and revised pins and circuitry after prototype testing exposed switching issues.
- Built a full-featured prototype with battery power, charging, USB-serial and Wi-Fi.
- Verified the PCB from an external designer, then handled manufacturing, SMD assembly and sourcing of components and connectors on a tight budget.
- Built the mat: layer construction, material sourcing, fabric printing and assembly.
- Designed and 3D-printed the enclosures and a webcam mount for foot imaging in Fusion 360.
- Wrote the signal processing, region mapping, rules engine, AI pipeline and PDF report generation.
Hardware
An ESP32 drives multiplexers that scan a flexible resistive sensor matrix. Each sensor sits in a voltage divider, and I tested resistor values to find the best fit for the range of pressures a standing person produces. Prototype testing exposed switching issues, so I revised the pin assignments and circuitry before the PCB was made.
Getting a clean signal
A resistive matrix shows ghosting: pressing a few points makes phantom readings appear where their rows and columns cross. I removed it with a median-filter-based algorithm that combines several signal operations.
For calibration I compared per-sensor slope calibration against normalised readings. For a non-medical-grade product, normalisation was enough, and it was much simpler to manufacture.
From readings to a heatmap
To save cost, the mat uses fewer, larger sensors. I mapped their readings onto foot regions using hand-drawn elliptical regions over a foot silhouette, then applied normalisation, colour mapping and a Gaussian blur to produce the pressure heatmap.
A guarded AI pipeline
Unconstrained LLM output wasn't reliable enough for health findings. So deterministic rules decide what the findings are, using a condition table I built from research into common foot conditions, with associated risks, exercises and insole recommendations. The model writes the personalised text and analyses static foot images alongside the sensor data. I upgraded it from GPT-4o mini to GPT-5.6 and tested the pipeline on a real dataset of 300 people.
Stack
- ESP32
- C++
- Multiplexers
- Python
- OpenCV
- GPT-5.6
- Fusion 360
- FDM printing