CAREC — Smart Wheelchair Safety

Open-source AI safety system for electric wheelchairs — real-time obstacle detection with directional alerts, built for children and designed to run entirely offline.

Built as a real embedded accessibility project combining edge AI, modular firmware, and caregiver-focused UX.

Accessibility / Edge AI / Assistive TechnologyActive development — Week 3

On-device AI — fully offline

Sub-200ms obstacle-to-alert

Free iOS & Android app

View Architecture Discuss a Similar Build

CAREC — Smart Wheelchair Safety system architecture and secure provisioning overview

Problem

Children in electric wheelchairs — especially younger users still building spatial awareness — can struggle to judge distance to obstacles in time to stop safely.

The safety system must run entirely on-device with no cloud dependency, produce alerts in under 200ms, survive hours of continuous use, and attach without modifying the original wheelchair.

System Architecture

The architecture is centred on a four-stage pipeline with clear module boundaries — motion gate, AI inference, distance estimation, and alert output — so each layer can be validated and extended independently.

Architecture Diagram Wiring & Relay Layout

Detection Pipeline & Alert Zones

Four-stage pipeline: motion gate → AI vision → distance assessment → alert decision, with three proximity zones (CRITICAL / WARNING / CLEAR) mapped to distinct beep patterns, display colours, and Bluetooth log events.

Three-Zone Proximity Model

CRITICAL (0–60 cm, fast beeps, red), WARNING (60–100 cm, slow beep, yellow), CLEAR (100 cm+, silence, green) — graded urgency without alert fatigue.

Mobile & Smart Home Integration

Free iOS/Android companion app for live status and event logs via Bluetooth; optional Home Assistant and Node-RED integration for caregiver notification dashboards.

Hardware & Software Stack

Detection & AI

On-device AI inference (no cloud)
Computer vision — object detection
Motion gating (optical flow, stationary suppression)
Bounding-box proximity heuristics
Confidence threshold filtering

Connectivity

Bluetooth 5 (mobile app, event logging)
WiFi 2.4 GHz (OTA firmware delivery)
MQTT (telemetry, Home Assistant topics)
Home Assistant + Node-RED integration

Alert & Interface

Three-zone directional audio (beep patterns)
Color-coded visual display (RED / YELLOW / GREEN)
iOS + Android companion app
Remote configuration via Bluetooth

Development & Delivery

C++17 — 7 modular header-only firmware modules
Python 3.11 — host tooling and integration tests
GitHub Actions CI/CD — build validation on every push
Unit + integration test suite
MIT open-source license

Detection approach	On-device AI object detection — no cloud
Alert latency	< 200ms from detection to audio output
Proximity zones	CRITICAL (0–60 cm) / WARNING (60–100 cm) / CLEAR (100 cm+)
Detection accuracy	85–92% in indoor environments
False positive rate	< 5% (motion gate + confidence threshold)
Battery runtime	50+ hours continuous operation
Mobile app	Free iOS + Android (Bluetooth, live status + logs)
Firmware updates	WiFi OTA — automatic, no cable required
Connectivity	Bluetooth 5 + WiFi 2.4 GHz
Mounting	Non-invasive tube clamp — removable in 2 minutes
License	MIT open source

How It Works

Motion gate filters idle frames

When the wheelchair is stationary, optical flow detects no motion and the inference pipeline is skipped — conserving processing budget and eliminating false positives from a static room.

On-device AI identifies objects

When the wheelchair moves, the AI accelerator runs object detection on the camera frame in real time. No cloud call, no internet required — inference completes in ~76ms on the local NPU.

Distance zone is classified

The apparent size of detected objects is used to estimate proximity. Objects are classified as CRITICAL (0–60 cm), WARNING (60–100 cm), or CLEAR (100 cm+) before the alert step.

Alert fires and event is logged

The appropriate beep pattern plays and the display updates within 200ms of detection. The zone event is logged over Bluetooth so the caregiver's mobile app stays current in real time.