peterindia.netSensor Data Fusion
How several sensors are combined into one better estimate: probabilistic filters, orientation and navigation methods, evidence-based reasoning, deep learning fusion, and the libraries that implement them.
Recursively combines noisy measurements and a motion model to estimate the state of a system, and is the workhorse of sensor fusion.
Best for: Linear systems with Gaussian noise.
Learn more →Applies the Kalman filter to non-linear systems by linearising around the current estimate.
Best for: Navigation and robot localisation.
Learn more →Uses the unscented transform to pass sample points through a non-linear model, often giving better accuracy than linearising.
Best for: Strongly non-linear models.
Learn more →Represents the estimate with many weighted samples, so it can handle non-linear, non-Gaussian problems.
Best for: Localisation and tracking with unusual noise.
Learn more →The general framework that updates a probability estimate as each new measurement arrives; Kalman and particle filters are special cases.
Best for: Understanding the theory behind the filters.
Learn more →Fuses estimates when the correlation between them is unknown, avoiding over-confident results.
Best for: Decentralised and multi-platform fusion.
Learn more →A simple fixed-gain tracking filter, a lightweight relative of the Kalman filter.
Best for: Small microcontrollers.
Learn more →Estimates states by optimising over a recent window of measurements, handling constraints directly.
Best for: Constrained systems.
Learn more →Models a hidden state that evolves over time and is seen only through noisy observations.
Best for: Activity and sequence recognition.
Learn more →Combines gyroscope, accelerometer and magnetometer data to estimate orientation, often with Madgwick or Mahony style filters.
Best for: Orientation of drones, phones and wearables.
Learn more →Dead-reckons position and velocity from inertial sensors, usually corrected by GNSS or other sensors.
Best for: Navigation without satellites.
Learn more →Estimates motion from a sequence of camera images, and is often fused with inertial data.
Best for: Robots and AR headsets.
Learn more →Builds a map of an unknown place while tracking the sensor's position within it, fusing cameras, LiDAR and inertial sensors.
Best for: Mobile robots and mapping.
Learn more →A graph formulation used to optimise a whole trajectory and map jointly from many sensor constraints.
Best for: Smoothing and optimisation-based fusion.
Learn more →Combines evidence from different sources and represents uncertainty and ignorance explicitly.
Best for: Conflicting or incomplete sensor evidence.
Learn more →Uses degrees of truth and rule-based reasoning to combine imprecise sensor information.
Best for: Control systems and rule-based fusion.
Learn more →A survey of datasets, methods and challenges in fusing camera, LiDAR and radar data with deep learning for driving.
Best for: Reading about fusion in autonomous driving.
Learn more →A survey of 3D object detection methods that fuse camera and LiDAR data in autonomous driving.
Best for: Comparing early, mid and late fusion in 3D detection.
Learn more →Open-source multi-task, multi-sensor fusion that unifies camera and LiDAR features in a bird's-eye-view space.
Best for: Camera and LiDAR fusion code.
Learn more →A large public autonomous-driving dataset with camera, LiDAR and radar data for training and testing fusion.
Best for: Benchmarking multi-sensor models.
Learn more →Widely used driving benchmarks with stereo, LiDAR and GPS/IMU data.
Best for: Evaluating fusion and odometry.
Learn more →Overview of combining sensory data from several sources to reduce uncertainty, including levels, architectures and examples.
Best for: The big picture.
Learn more →Background on integrating data from multiple sources, covering the wider field beyond sensors.
Best for: Related terminology and methods.
Learn more →MathWorks toolbox with algorithms and examples for fusing sensor data and tracking objects.
Best for: Prototyping and simulation.
Learn more →ROS package that fuses odometry, IMU, GPS and other sensors into a state estimate using Kalman filters.
Best for: Mobile robots on ROS.
Learn more →C++ library for smoothing and mapping using factor graphs, with Python bindings.
Best for: Optimisation-based fusion.
Learn more →Python library for Kalman filters and other optimal estimators.
Best for: Quick Python experiments.
Learn more →An open textbook in Jupyter notebooks that teaches Kalman and Bayesian filtering step by step.
Best for: Learning the maths with code.
Learn more →An illustrated walkthrough of the Kalman filter's ideas and equations.
Best for: A gentle first read.
Learn more →A small C library for IMU and AHRS sensor fusion that runs on embedded devices.
Best for: Orientation on microcontrollers.
Learn more →PX4's documentation for its navigation filter, which fuses IMU, GPS, barometer and other sensors.
Best for: Drone state estimation.
Learn more →ArduPilot's overview and tuning guide for its extended Kalman filter navigation system.
Best for: Drone and vehicle navigation.
Learn more →Match your requirement to a starting point, then compare vendors.
| Learn the basics | Start with the Kalman filter and the illustrated tutorial |
|---|---|
| Non-linear system | Extended or unscented Kalman filter, or a particle filter |
| Phone, wearable or drone orientation | An AHRS filter, such as the x-io Fusion library |
| Robot localisation and mapping | SLAM, factor graphs and robot_localization |
| Self-driving perception | Camera and LiDAR fusion such as BEVFusion, with nuScenes or KITTI |
| Uncertain or conflicting evidence | Dempster–Shafer theory or fuzzy logic |
| Prototype quickly | MATLAB toolbox or FilterPy |
Neighbouring directories and hubs.
Quick answers.
It is the process of combining data from several sensors so that the result is more accurate, complete or reliable than any single sensor could give.
Each sensor has noise, drift or blind spots. For example, gyroscopes drift over time and accelerometers are noisy, but combining them gives stable orientation.
Data can be combined as raw signals (early), as extracted features (mid) or as decisions (late). Deep learning systems often compare these choices.
A Kalman filter assumes simple, Gaussian noise and is very efficient. A particle filter uses many samples and copes with non-linear and non-Gaussian problems, at a higher computing cost.
For most projects start with a Kalman filter or an extended Kalman filter, since there are many tutorials and libraries.
Algorithm cards link to Wikipedia articles, and tools and papers link to their own pages. Links were checked on 7 October 2026 and descriptions are written for this page.
Algorithm descriptions are written for this page and link to Wikipedia or to each tool's own page (checked 7 October 2026).