An accelerometer measures linear acceleration and can infer tilt from gravity; a gyroscope tracks angular velocity; an IMU packages both (and sometimes a magnetometer) into one device. The right choice depends on whether you need to measure impact, rotation, or full orientation - and whether you can afford the drift, noise, and calibration overhead that comes with each.
Key Takeaways:
- Accelerometers work well for tilt under near-static conditions but break down when dynamic forces dominate.
- Gyroscopes track fast rotation cleanly but drift relentlessly when you integrate rate into angle - bias errors compound every sample.
- A 6-axis IMU doesn't magically solve orientation; you still need fusion algorithms, calibration discipline, and realistic expectations about long-term accuracy.
- Adding a magnetometer (9-axis) helps with heading only if your environment isn't full of motors, steel, or other magnetic noise.
- Specification sheets lie by omission - bias stability, temperature drift, and cross-axis coupling matter more than the number of bits in the ADC.
What Each Sensor Actually Measures
An accelerometer outputs specific force along three orthogonal axes (X, Y, Z), typically in units of g or m/s². It responds to both applied acceleration and gravity. A stationary accelerometer on a table doesn't read zero - it reads approximately 1 g on the axis perpendicular to the surface because the table is pushing back against gravity.
This gravity response is what allows tilt estimation. Rotate the sensor, and the 1 g vector redistributes across the three axes. Simple trigonometry extracts pitch and roll angles.
But here's the catch: the math assumes gravity is the only acceleration present. Shake the sensor, mount it on a moving vehicle, or subject it to vibration, and those dynamic forces corrupt the tilt estimate. I've debugged systems where a drone's accelerometer reported wild pitch changes during aggressive climbs - not because the IMU failed, but because forward acceleration overwhelmed the gravity signal the algorithm expected.
Accelerometers also can't sense rotation around the gravity vector (yaw for a level device). Spin a flat sensor on a table and the accelerometer sees almost nothing.
A gyroscope measures angular velocity - how fast something rotates around an axis - in degrees per second (°/s) or radians per second. MEMS gyroscopes provide three channels corresponding to roll rate, pitch rate, and yaw rate, depending on your body-frame convention.
Gyroscopes don't output absolute angle. To recover orientation, you integrate angular velocity over time:
θ(t) = θ₀ + ∫₀ᵗ ω(τ) dτ
where θ(t) is the angle estimate, θ₀ is the initial angle, and ω is the measured rate.
Integration amplifies every error. A gyroscope with 0.1 °/s of residual bias - common even after calibration - accumulates 6° of angle error per minute. Real gyroscopes exhibit bias instability, temperature-dependent drift, and noise, so uncorrected integration becomes unusable within seconds to minutes depending on the sensor grade.
An IMU is a module that combines inertial sensors into one package. A typical 6-axis IMU contains a three-axis accelerometer and a three-axis gyroscope, providing six measurement channels. Many modern IMUs add programmable ranges, digital filtering, FIFO buffers, motion-triggered interrupts, and even embedded processing. STMicroelectronics' LSM6DSV, for example, integrates both sensor types with configurable full-scale ranges (±2 to ±16 g for accel, ±125 to ±4000 °/s for gyro), onboard filtering, and sensor-fusion support.
An IMU does not automatically deliver accurate roll, pitch, or yaw. Those require calibration, fusion algorithms, and often external references. The IMU gives you raw data; everything else is your job.
Why Sensor Fusion Exists
Accelerometers provide a long-term gravity reference for pitch and roll but suffer from dynamic interference. Gyroscopes track fast rotation cleanly but drift when integrated. Sensor fusion combines both to exploit their complementary strengths.
A complementary filter is one of the simplest fusion methods:
θₑₛₜ = α(θₚᵣₑᵥ + ωΔt) + (1 - α)θₐcc
where θₑₛₜ is the updated angle estimate, α weights the gyroscope contribution (typically 0.95-0.98), θₚᵣₑᵥ is the previous estimate, ω is angular velocity, Δt is the sample interval, and θₐcc is the accelerometer-derived tilt.
The gyroscope dominates short-term response, while the accelerometer gently corrects long-term drift.
In practice, I've found that tuning α is less about theory and more about trial: too high and the estimate drifts; too low and vibration injects noise into what should be smooth rotation. The optimal value depends on your sensor quality, motion profile, and how much you trust the accelerometer under your specific operating conditions.
More sophisticated systems use Kalman filters, extended Kalman filters (EKF), or Madgwick algorithms. These explicitly model sensor noise and dynamics but cost more CPU and require more tuning. A drone flight controller running at 1 kHz might justify an EKF; a low-power wearable sampling at 50 Hz probably doesn't.
When a Magnetometer Helps (and When It Doesn't)
A magnetometer measures the local magnetic field vector across three axes. When fused with accelerometer and gyroscope data, it can provide a heading reference relative to Earth's magnetic field - hence the term "9-axis" system.
Bosch's BNO055 is a common example: it integrates a three-axis accelerometer, gyroscope, magnetometer, and an ARM Cortex-M0 running proprietary fusion firmware that outputs quaternion or Euler-angle orientation.
But magnetometers are fragile references. Motors, magnets, steel structures, current-carrying wires, speakers, and even nearby laptops distort the local field. I've seen indoor robots where the magnetic heading wandered ±30° depending on which room the robot occupied, purely due to building steel and HVAC equipment.
If your application operates in a magnetically clean environment and you need absolute heading without GPS, a 9-axis setup makes sense. If you're near motors, in a vehicle, or indoors with unknown magnetic conditions, the magnetometer often adds more calibration burden than value. In those cases, a 6-axis IMU plus an external heading source (GPS, wheel odometry, visual landmarks) is usually more reliable.
The Specifications That Actually Matter
Full-Scale Range
Accelerometer ranges are typically ±2, ±4, ±8, or ±16 g. Gyroscope ranges run from ±125 to ±2000 °/s or higher. The ST LSM6DSV offers ranges up to ±16 g and ±4000 °/s.
Wider isn't always better. A ±16 g accelerometer has coarser least-significant-bit resolution than a ±2 g setting (assuming the same ADC bit depth), and noise specs often degrade at higher ranges. Choose the smallest range that won't saturate during normal operation plus transients.
A tilt sensor sitting on a desk needs ±2 g. A quadcopter experiencing aggressive maneuvers might need ±8 g. An impact recorder in a shipping container might need ±16 g or more.
Noise Density and Bias Stability
Noise density (typically µg/√Hz for accelerometers, m°/s/√Hz for gyroscopes) tells you how much measurement noise appears per unit bandwidth. Lower is better, but you also need to consider your application bandwidth and filtering.
Bias stability describes how much the zero-rate offset wanders over time. For gyroscopes, this is critical: a gyro with 1 °/s bias instability will drift 60° per minute if you integrate without correction. High-grade IMUs (tactical or navigation-grade, often $100-$1000+) achieve bias stability below 1 °/hr. Consumer MEMS gyros (under $10) are closer to 10-50 °/hr or worse.
I've learned not to trust bias specs blindly. Measure the actual bias behavior in your system, at operating temperature, after your reflow profile. Datasheets show typical or maximum values under controlled conditions that don't always match the real world.
Temperature Drift
Bias changes with temperature. A gyroscope might show 0.05 °/s bias at 25°C and 0.3 °/s at 60°C. If your system experiences thermal cycles - inside a car dashboard, near a motor, in sunlight - temperature compensation becomes mandatory.
Some IMUs include internal temperature sensors and factory-calibrated compensation coefficients. Others require you to characterize drift yourself and apply corrections in software.
Sampling Rate and Latency
Output data rate (ODR) determines how often you get new samples. Bandwidth determines what frequency content those samples represent.
A 1 kHz ODR sounds impressive, but it's overkill if your mechanical system can't change faster than 50 Hz and you're bandwidth-limited by a 100 Hz digital filter inside the IMU. Faster sampling increases I²C/SPI traffic, interrupt overhead, and power consumption without improving measurement quality.
Match ODR and bandwidth to your control-loop rate and the fastest dynamics you need to observe. A drone might run IMU sampling at 1 kHz to feed an 800 Hz attitude controller. A fitness tracker logging steps once per second can get by with 50 Hz or less.
6-Axis vs. 9-Axis: What You Actually Get
| Factor | 6-Axis IMU | 9-Axis System |
|---|---|---|
| Accelerometer | 3 axes | 3 axes |
| Gyroscope | 3 axes | 3 axes |
| Magnetometer | None | 3 axes |
| Gravity-based tilt | Yes | Yes |
| Absolute heading (magnetic) | No | Yes, if environment permits |
| Magnetic interference concern | Not a sensor issue | Can corrupt heading |
| Calibration complexity | Accel + gyro | Accel + gyro + mag (3D ellipsoid fit) |
| Typical use case | Stabilization, motion tracking without heading | Orientation with heading reference |
The "9-axis" label refers to nine sensor channels, not nine mechanical degrees of freedom. A rigid body in 3D space has six DOF (three translations, three rotations). You're measuring those six DOF using nine data streams, some of which are redundant or complementary.
Real-World Application Tradeoffs
Drones demand low latency, high sample rates, and good gyroscope performance. Vibration from motors is a persistent problem - mechanical isolation (rubber standoffs, foam) and notch filters tuned to motor frequencies help. Magnetic heading is often unreliable near brushless motors, so GPS velocity or visual odometry provides yaw reference instead.
Wearables prioritize power, size, and embedded motion detection over raw accuracy. Many integrate activity-recognition algorithms directly on the IMU to reduce MCU wake-ups. Absolute orientation isn't critical; detecting steps, gestures, or sleep posture is.
Robotics (balancing robots, gimbals, exoskeletons) need responsive angular-rate feedback and reasonable tilt accuracy. Sensor-to-body alignment matters - a 2° mounting misalignment directly biases your roll/pitch estimates. I've spent hours debugging "mysterious lean" in a two-wheeled robot that turned out to be a 1.5° IMU tilt on the PCB.
Smartphones use IMUs for screen rotation, gaming, and augmented reality. They've driven down the cost and size of consumer MEMS sensors but also established expectations for "good enough" accuracy - users tolerate a few degrees of drift in AR apps because the experience resets frequently.
Industrial machinery monitoring uses accelerometers for vibration analysis (bearing wear, imbalance) and inclinometers for tilt or leveling. Here, the concern is often harsh environments (temperature, shock, EMI) rather than high-speed control loops.
Selecting the Right Sensor for Your Design
Start with the Motion, Not the Sensor
What are you trying to measure? Tilt? Rotation rate? Full 3D orientation? Impact detection? Vibration frequency?
If you only need to know "is the device tilted more than 15°," a single-axis accelerometer might suffice. If you need to detect rotation events but don't care about absolute angle, a gyroscope alone works. If you need stable pitch/roll/yaw over minutes, you need a 6-axis IMU with fusion.
Define Operating Conditions
- Motion range: What's the maximum expected acceleration and rotation rate, including transients?
- Environment: Temperature range, vibration spectrum, magnetic interference, mounting constraints.
- Power budget: Always-on sampling vs. motion-triggered wake-up.
- Latency tolerance: Control loops need fresh data; data logging can tolerate delays.
Evaluate Sensor Specs in Context
Don't compare resolution alone. A 16-bit ADC doesn't help if noise occupies the bottom 4 bits. Look at:
- Noise density at your chosen range and bandwidth
- Zero-g offset and zero-rate bias (and their stability)
- Temperature coefficients
- Cross-axis sensitivity (ideally <2%)
- Supply voltage and current consumption
Decide Where Fusion Happens
Some IMUs (like the LSM6DSV or BNO055) include embedded processing that outputs orientation directly. Others provide raw data and expect your MCU to run the fusion algorithm.
Embedded fusion offloads your processor and can reduce power (the IMU's coprocessor is often more efficient than waking your main MCU at high rate). But you sacrifice configurability - you're stuck with the vendor's algorithm and update rate.
If you need tight integration with other sensors (cameras, LiDAR, encoders) or custom fusion logic, raw-data IMUs give you control at the cost of software complexity.
Verify Interface and Integration
I²C is convenient for multi-sensor buses and simple wiring but limited to ~400 kHz (or 1 MHz in fast mode+). SPI is faster and more deterministic, at the cost of more pins. Some newer IMUs support I3C for higher speed and lower power.
Check that your MCU supports the interface speed and timing you need. A 1 kHz ODR over I²C at 400 kHz is feasible for a small packet (maybe 14 bytes), but marginal if you're also polling a magnetometer, barometer, and other devices on the same bus.
Common Problems and How to Fix Them
Orientation drifts slowly even when stationary: Gyroscope bias. Collect a few seconds of stationary data at startup, average the readings, and subtract that offset from subsequent samples. This only corrects constant bias - instability and temperature drift remain.
Tilt angle changes when you accelerate the device: Dynamic acceleration is mixing with gravity in the accelerometer output. Increase the time constant (lower 1 - α) in your complementary filter to trust the gyroscope more during motion, or add acceleration-magnitude monitoring to detect when the accelerometer is unreliable.
IMU readings are excessively noisy: Check mechanical mounting (loose screws transmit vibration), review bandwidth and filter settings in the IMU registers, and inspect your power supply for noise. I've traced "noisy IMU" issues to inadequate decoupling capacitors on the sensor's VDD pin.
Motion saturates at a fixed value: You've hit the full-scale limit. Increase the range. Then revisit whether you really need to measure that extreme motion or whether the event is an anomaly you should detect separately (e.g., drop detection with a dedicated high-g threshold).
Axes are backwards or swapped: Coordinate frame mismatch. Draw the sensor orientation on your PCB, confirm the X/Y/Z silkscreen or datasheet body diagram, and map it explicitly to your vehicle/robot/device frame. Sign errors here will make your control system fight itself.
Magnetic heading spins randomly: Nearby motors, magnets, or steel. Move the magnetometer if possible, improve calibration, or disable magnetic heading and use GPS/odometry/other sources.
Timing jitter causes angle jumps: If you integrate gyroscope rate with a nominal Δt but actual sample intervals vary by milliseconds, angle errors accumulate. Timestamp every sample (from the IMU's internal counter or your RTOS tick) and use the actual interval in your integration.
What This Means for Your Next Design
For tilt sensing in near-static conditions - think construction levels, platform leveling, or orientation detection in consumer devices - an accelerometer alone often works and costs less. Accept that motion and vibration will disturb it.
For rotation tracking without long-term angle accuracy - rate gyros in camera gimbals, turntable speed sensing - use a gyroscope. Don't integrate to angle unless you reset or correct frequently.
For short-duration orientation (seconds to minutes) in dynamic environments - drone stabilization, VR headsets, handheld gimbals - a 6-axis IMU with complementary or Kalman filtering delivers practical results. Calibrate gyro bias at startup and tune your filter to match your motion profile.
For long-duration navigation or orientation over minutes to hours, fuse the IMU with external references: GPS velocity, wheel encoders, visual odometry, or fixed landmarks. The IMU provides high-rate motion updates between slower absolute measurements.
The sensor is not the system. The mechanical mount, thermal environment, calibration process, fusion algorithm, and external references determine whether you get usable orientation or a drifting mess. Choose the sensor that matches your motion range and error budget, then invest the effort to integrate it properly.
Frequently Asked Questions
Can an accelerometer measure rotation?
No. It measures linear acceleration and can infer tilt from the gravity vector under near-static conditions, but it does not measure angular velocity or rotation around the gravity axis.
Why combine an accelerometer and gyroscope?
They complement each other. The gyroscope tracks fast rotation but drifts when integrated. The accelerometer provides a long-term gravity reference for pitch and roll but breaks down during dynamic motion. Fusion corrects each sensor's weakness.
What does "6-axis" or "9-axis" mean?
A 6-axis IMU has three accelerometer axes and three gyroscope axes. A 9-axis system adds three magnetometer axes. These are sensor channels, not mechanical degrees of freedom.
Does an IMU give me position without GPS?
Only briefly. Double-integrating accelerometer data accumulates error so fast that position estimates become unusable within seconds unless you have external corrections (wheel encoders, visual odometry, etc.).
Why does my calibrated IMU still drift?
Calibration reduces bias and scale errors but doesn't eliminate noise, temperature effects, or time-varying instability. Gyroscope integration amplifies every remaining error. Sensor fusion and external references are required for long-term stability.
When should I add a magnetometer?
When you need absolute heading, operate in a magnetically clean environment, and can handle the calibration complexity (3D ellipsoid fit, often requiring the user to wave the device in a figure-8 pattern). If motors, steel, or magnets are nearby, the magnetometer usually causes more trouble than it solves.
