SmarterCanes

Low-Cost Sensing for Step Counting and Self-Supervised Gait Assessment

A low-cost instrumented walking cane — an IMU and a downward-facing time-of-flight distance sensor, logging to SD — paired with an analysis pipeline that counts steps and flags abnormal gait from an autoencoder trained only on normal walking.

The project has three parts:

• Firmware — an ESP32-based recorder (accelerometer, gyroscope, distance sensor) with button-driven session control and a WiFi dashboard for offloading data. See also Firmware/POWER_ANALYSIS.md for the current-draw budget and estimated battery life.
• Dataset — the recorded IMU sessions and their labels (train/test/walk split, per-recording annotation, anomaly type).
• Analyzer — the Python pipeline that consumes the dataset: a step counter scored against the distance sensor’s ground truth, and a self-supervised (one-class) autoencoder that learns normal gait and scores deviations from it. See Analyzer/EXPERIMENT.md for the full experimental design and rationale behind the anomaly-detection pipeline.

How the pieces fit together

The cane (Firmware) records 6-axis IMU data and cane-tip height to SD as CSV. Recordings are collected into the Dataset and split into train (normal gait only) and test (labelled, includes abnormal gait). The Analyzer’s step counter turns raw acceleration into a step count, cross-checked against the distance sensor; its autoencoder is fitted on train and evaluated frozen on test, so no anomaly label ever enters the pipeline that produces the score meant to catch it.

Getting started

• To build and flash the firmware: see Firmware/README.md.
• To run the analysis on the existing dataset: see Analyzer/README.md (cd Analyzer && uv sync, then uv run main.py ...).