Skip to main content

wearable-activity-fall

:::warning [UNTESTED] — v0.9 paper-correct The two cores (motion_features, fall_detect) are host-unit-tested on native_sim/native/64; the full app runs end-to-end on native_sim with synthetic motion (incl. an injected fall). No model ships (stub + deterministic fallback); the fall detector needs no model. HiL on a real wearable + a trained model is bench-gated. :::

Body-worn IMU edge node: read a 6-axis IMU, classify coarse activity (idle/walk/run/stairs) with a small NPU model, and detect falls with a rule-based 3-phase detector. Targets wearables / elder-care / lone-worker safety.

Source: examples/ai/wearable-activity-fall/.

Honest scope​

Body-worn motion sensing. Detects falls + coarse activity. Not medical-grade, not a certified fall alarm, no gait/health diagnostics. The fall detector is a physics heuristic (tunable thresholds: free-fall → impact → stillness), not a guarantee.

Pipeline​

ICM-42670 accel+gyro (I2C, 100 Hz, +/-16 g)
  | every sample -> fall_detect (free-fall -> impact -> stillness)
  | 256-sample window -> motion_features (RMS/SMA/cadence/jerk/tilt)
  |   -> <alp/inference.h> activity classifier (deterministic fallback)
  -> WACT record per window

The IMU runs at ±16 g so fall impacts (several g) do not clip.

board.yaml​

som:
  sku: E1M-AEN701

preset: e1m-evk
supported_boards:
  - e1m-evk
  - e1m-x-evk

pins:
  - { e1m: E1M_I2C0, macro: EVK_I2C_BUS_SENSORS, doc: "ICM-42670 IMU bus" }

cores:
  a32_cluster:
    os: "off"
  m55_hp:
    app: ./src
    inference:
      default_arena_kib: 64
    libraries:
      - tflite_micro
    peripherals:
      - i2c                   # ICM-42670 link

chips:
  - icm42670                  # 6-axis IMU; accel + gyro

diagnostics:
  log_level: info

Flip som.sku to E1M-V2M101 for the DEEPX DX-M1 path.

Expected output​

# WACT,t_s,activity,confidence,fall,impact_g
WACT,2.56,WALK,0.91,0,0.0
WACT,12.80,IDLE,0.74,1,4.8

See also​

Questions about this page? Discuss in Community Forum