← AI Terminology
Kalman Filter
The Kalman filter is a recursive Bayesian algorithm that estimates the true state of a dynamic system from noisy measurements — combining a prediction step (using a motion model) with an update step (incorporating new observations) to maintain an optimal running estimate.
It is used wherever a sensor is noisy and a model of motion is available.
It is used wherever a sensor is noisy and a model of motion is available.
Why It Matters in AI
Kalman filters are the backbone of sensor fusion in robotics, autonomous vehicles, and aerospace — combining GPS, IMU, LIDAR, and camera data into a single consistent state estimate. They are optimal (minimum variance) for linear Gaussian systems, and Extended/Unscented variants handle nonlinear dynamics. In AI, they appear in tracking (object tracking in video), SLAM (simultaneous localisation and mapping), and time series smoothing.
Key Points
| Aspect | Description |
|---|---|
| EKF | Extended Kalman Filter — linearises nonlinear dynamics via Jacobians |
| UKF | Unscented Kalman Filter — uses sigma points instead of linearisation; more accurate for nonlinear |
| Two steps | Predict: propagate state estimate forward using motion model; Update: correct with new measurement |
| Optimality | Minimum mean-squared-error estimator for linear systems with Gaussian noise |
| Applications | Object tracking, SLAM, drone navigation, financial forecasting, speech enhancement |
| Particle filter | Non-parametric alternative for highly nonlinear/non-Gaussian systems; computationally heavier |
Simple Analogy
A ship's navigator using dead reckoning: predict position from speed and heading (motion model), then correct with a GPS fix when available (measurement update). The Kalman filter does this optimally — weighting the GPS fix by how noisy it is vs. how confident the dead-reckoning estimate is.
Common Usage Examples
filterpy.kalman.KalmanFilter— Python Kalman filter library used in tracking applications- SORT / DeepSORT: Kalman filter tracks bounding box motion between frames in multi-object tracking
- ROS
robot_localizationpackage: EKF fusing GPS + IMU for mobile robot state estimation - Autonomous vehicles: sensor fusion for lane-keeping using EKF over LiDAR + camera + wheel odometry
- Financial:
pykalman.KalmanFilterfor smoothing noisy price time series
Summary
In short: The Kalman filter is the optimal algorithm for fusing noisy sensor measurements with motion predictions — the workhorse of robotics, autonomous vehicles, and any system where tracking a moving state matters.