Sensors and Sensor Fusion: Estimating Reality from Noisy Measurements

A sensor does not report reality directly. It converts a physical property into an electrical or digital measurement with uncertainty.

Sensor fusion combines several imperfect measurements and a model over time to estimate a state more reliably than one sensor can alone.

The estimate might describe position, orientation, speed, temperature, occupancy, or an object's identity.

A concrete example: phone navigation

A phone combines:

• GPS for global position,
• accelerometer for linear motion,
• gyroscope for rotation,
• magnetometer for heading,
• barometer for elevation changes,
• and map constraints.

GPS may disappear indoors. Inertial sensors drift. Together, the system can maintain a smoother estimate and correct errors when stronger evidence returns.

Measurement and state

The physical state is what the system wants to know. A measurement is what a sensor observes.

A wheel encoder measures rotation, not vehicle location. A camera measures light, not object identity. A thermometer measures its own sensing element, which may lag the surrounding air.

A measurement model connects observations to the hidden state.

Noise

Noise creates random variation around a value.

Sources include electronics, vibration, light, radio interference, manufacturing variation, and quantization. Repeated readings under apparently fixed conditions show a distribution rather than one perfect number.

Filtering can reduce noise but may also add delay or hide rapid real changes.

Bias

Bias is a systematic offset.

A scale may read 200 grams too high. A gyroscope may report slight rotation while stationary. Averaging does not remove stable bias.

Calibration estimates bias against a known reference, and systems may update calibration as temperature or age changes.

Drift

Drift is change in sensor behaviour over time.

It can come from temperature, wear, contamination, component aging, or accumulated integration error. Inertial position estimates drift because tiny acceleration errors accumulate into velocity and position.

Fusion uses independent references to correct drift.

Resolution and range

Resolution is the smallest distinguishable change. Range is the span the sensor can measure.

A high-resolution sensor may saturate outside a narrow range. A wide-range sensor may miss small changes. Some systems combine sensors with different ranges for ordinary and extreme conditions.

Sampling

Sampling rate determines how often measurements are taken.

Too slow can miss fast motion or misrepresent frequency. Faster sampling consumes energy, storage, and computation and may collect mostly noise.

Filters and control loops need sampling schedules matched to the dynamics being observed.

Time synchronization

Fusion assumes measurements refer to known times.

Camera, GPS, and inertial sensors may have different clocks, delays, and rates. A 50-millisecond timestamp error can create a large spatial mismatch on a moving vehicle.

Use synchronized clocks, hardware timestamps where needed, and models that account for latency.

Coordinate frames

Measurements may be expressed in:

• sensor frame,
• device frame,
• vehicle frame,
• room frame,
• or global geographic frame.

Fusion requires known transformations among them. Incorrect axis direction or mounting orientation can look like mysterious model failure.

Complementary sensors

Sensors are valuable when their errors differ.

A gyroscope captures quick rotation smoothly but drifts. A magnetometer provides a long-term heading reference but is disturbed by nearby metal. Combining them can preserve rapid response while correcting gradual error.

Two identical sensors with the same failure mode provide less independent information.

Filtering

Simple filters include moving averages and low-pass filters. More advanced estimators include complementary filters, Kalman-filter families, particle filters, and optimization-based methods.

The method represents assumptions about motion, measurement noise, and uncertainty. Complexity is justified only when it improves the target estimate under real conditions.

Kalman-filter intuition

A Kalman filter alternates:

1. predict the next state from the previous state and motion model,
2. compare predicted measurements with sensor observations,
3. and update the estimate according to uncertainty.

More trusted measurements influence the update more strongly. Nonlinear systems use extensions or other methods.

Outliers and sensor failure

Measurements can be impossible rather than merely noisy.

Detect:

• values outside physical limits,
• frozen readings,
• sudden jumps,
• inconsistent sensors,
• missing samples,
• and degraded signal quality.

Reject, down-weight, or enter a safe degraded mode. Fusion should not make one failed sensor appear confidently correct.

Uncertainty

A useful estimator reports uncertainty as well as a best estimate.

Downstream systems can slow a robot, ask for user confirmation, widen a safety margin, or stop when uncertainty is too high.

Confidence must reflect real conditions and be evaluated for calibration.

Training-data concerns

Machine-learning-based perception depends on labelled examples covering lighting, weather, devices, users, environments, and rare events.

Evaluate by condition and sensor revision. A model may fail after a camera supplier changes even when the interface specification remains the same.

Calibration and maintenance

Record calibration procedure, date, hardware identity, software version, and environmental conditions.

Provide field diagnostics and replacement rules. A sensor system is maintained physical equipment, not a one-time algorithm deployment.

Validate against independent ground truth

Compare estimates with a more trusted reference across the full operating range, including transitions and failures. Report error distributions by environment rather than one average.

Replay recorded synchronized sensor data to reproduce regressions, but confirm fixes on physical hardware because timing, vibration, temperature, and mounting may not replay faithfully.

Knowledge check

1. How do noise, bias, and drift differ?
2. Why is timestamp alignment important?
3. What makes two sensors complementary?
4. How does prediction-and-update fusion work?
5. Why should an estimator expose uncertainty?

The one idea to remember

Sensors provide imperfect observations, not direct truth. Reliable sensing models noise, bias, drift, timing, frames, and failure, then combines complementary evidence over time while exposing uncertainty and maintaining calibration.