Sensor fusion is the combination of measurements from multiple sensors into a single estimate that exploits each sensor’s strengths and corrects against its weaknesses. When the sensors’ failure modes are largely independent, the fused estimate is more accurate than any one alone — though a poorly tuned fusion can be less accurate than a trusted single sensor if it overweights a noisy one. The classic example is the IMU: an Accelerometer senses gravity, a Gyroscope senses rotation rate, and a Magnetometer senses magnetic north, and fusing all three gives a stable estimate of orientation in space.
The need for fusion comes from each sensor’s individual flaws. The gyroscope, if used alone, drifts — its integrated orientation diverges from truth over minutes. The accelerometer alone confuses gravity with motion-induced acceleration when the device is moving. The magnetometer alone is easily fooled by local magnetic disturbances. Each sensor has a different failure mode, and the failures are largely independent. Combining them — using each sensor where it’s strongest, and using the others to correct its weaknesses — produces an estimate that none of them could produce alone.
The algorithmic core of sensor fusion is usually a state-space filter. The Kalman filter is the textbook choice — it maintains a running estimate of orientation along with an uncertainty, predicts forward using the gyroscope readings, and corrects against the accelerometer and magnetometer with weights determined by each sensor’s noise characteristics. For nonlinear orientation dynamics the Extended Kalman filter (EKF) is standard. Simpler and cheaper alternatives like the Complementary filter (a high-pass on the gyro-integrated angle plus a low-pass on the accelerometer-derived tilt) and the Madgwick filter (a gradient-descent quaternion update) are common on resource-constrained hardware and give surprisingly competitive accuracy.
The fusion is what makes a phone know whether you’re holding it in landscape or portrait, makes a drone hover stably, makes a VR headset track head motion without lag. All of these are orientation estimates that no single sensor could produce.