Almost every wearable ships with an inertial measurement unit, and almost every wearable team picks the wrong one first. The datasheet columns that look important — resolution in bits, maximum sample rate, full-scale range — rarely decide whether step counting works on a wrist. The specs that decide it are noise density, zero-g offset drift over temperature, and how much current the part draws in the mode you will actually run it in.

An IMU costs between $1.20 and $6.00 in volume. Choosing badly does not cost you much on the bill of materials. It costs you six months of algorithm work trying to filter out a problem the sensor created.

What an IMU Contains and What You Actually Need

A six-axis IMU packages a three-axis MEMS accelerometer and a three-axis MEMS gyroscope in one 2.5 mm by 3 mm LGA. A nine-axis part adds a magnetometer, which gives absolute heading and is ruined by any nearby magnet, motor, speaker, or ferrous fastener — a description of most of the inside of a wearable. Unless you have a real compass requirement and a plan for hard-iron calibration on every unit, start with six axes.

Many activity products need only the accelerometer. Step counting, sleep staging, wear detection, orientation, and impact detection all work from acceleration alone, at a tenth to a hundredth of the gyro's current. Gyros earn their power budget when you need short-term angular rate: gesture recognition, range-of-motion measurement in rehab devices, or gait analysis.

The Specs That Change Behavior

Accelerometer

  • Noise density, in micro-g per root hertz. Under 100 is good for a consumer wearable; under 60 matters for sleep and tremor work. This number, not bit depth, sets usable resolution.
  • Full-scale range. Wrist motion rarely exceeds 8 g, but an impact against a table hits 16 g and clips. Pick a part that switches between a low range for fine motion and 16 g for impact detection.
  • Zero-g offset and temperature drift. A device worn against skin sees 20 to 25 degrees C of swing. A part drifting 1 mg per degree C moves its baseline enough to corrupt tilt-based posture estimates.
  • Output data rate and FIFO depth. A 512-sample FIFO lets the MCU sleep for seconds and wake once to drain a burst. Without it, the MCU wakes on every sample and the power budget collapses.

Gyroscope

  • Zero-rate offset and drift is the dominant gyro error. Any angle computed by integrating rate accumulates it linearly, which is why a gyro alone can never hold absolute orientation.
  • Full-scale range is typically selectable up to ±2,000 degrees per second. A fast wrist flick reaches 800; a golf swing exceeds that.
  • Vibration rectification. MEMS gyros convert broadband vibration into a DC bias. If the product sits near a haptic motor or a pump, test this before committing.

Where on the Body It Goes

Placement changes the signal more than any datasheet spec. The wrist is convenient and noisy: hand gestures unrelated to locomotion swamp step signals. The chest and upper back sit near the body's center of mass and give the cleanest posture, respiration, and gait data, which is why clinical instruments go there. The waist is the classic pedometer location and the most accurate for step counting. The ankle gives excellent gait phase detection and terrible everything else.

Inside the enclosure, mount the IMU rigidly to the PCB and the PCB rigidly to the housing. Any compliance in that chain — foam, a floating board on long standoffs — becomes a mechanical low-pass filter with a resonance you did not design. Keep it away from speakers, haptic actuators, and pulsed battery contacts, and note the sensor's axis orientation on the drawing so it stays consistent across revisions. The mounting rules are those in mounting a PCB in an enclosure; comfort and fit belong in wearable product design.

The Power Budget

Take a 60 mAh lithium polymer cell in a band and a target of seven days between charges. That is about 350 microamps average for everything: sensor, MCU, radio, display.

A modern accelerometer at 25 Hz in low-power mode draws 5 to 15 microamps; the same part at 200 Hz in high-performance mode draws 150. A gyro draws 500 to 900 microamps whenever it is on, with no low-power mode worth the name, so running one continuously is not compatible with a week of battery life on a small cell.

The pattern that works is duty cycling by event: keep the accelerometer at low rate with its interrupt engine armed for motion, tap, or free-fall, let the MCU sleep, and when activity fires, wake, turn on the gyro for the two seconds the gesture lasts, then shut it off. That discipline is covered in low-power firmware and sleep modes, and the cell sizing math in battery pack design for a product.

Sensor Fusion: Where the Math Runs

Fusion combines the accelerometer's stable long-term gravity reference with the gyro's clean short-term rate to produce an orientation neither gives alone, and you have three places to run it.

On the IMU. Several parts embed a fusion core and output quaternions directly: fast to integrate, minimal MCU load, and a black box you cannot tune when it misbehaves on your motion.

On your MCU. A complementary filter is forty lines of code and works for tilt. A Madgwick or Mahony filter costs a few percent of a Cortex-M4 at 100 Hz and handles full orientation. This is the default for most products, and it is why microcontroller selection should account for a floating-point unit.

Off-device. Streaming raw data to a phone is cheap on hardware, expensive on radio power, and useless offline. The middle path is a small trained model on the MCU, described in TinyML on microcontrollers.

Validate on Bodies, Not on a Bench

Bench rotation on a fixture proves the part is alive and nothing about the product. Collect labeled raw data from at least twenty people across body types, wear positions, and activities, and keep that dataset — it becomes the regression suite for every future firmware change. Start it before the enclosure is frozen, because the data usually moves the sensor.

Getting the Sensor Choice Right the First Time

Projects House selects and characterizes inertial sensing for wearable products end to end: part shortlist against your actual motion signature, placement studies, power modeling against a real cell, and the fusion stack. Describe the motion you need to measure and your battery target through the contact form.