Sensor Fusion: Classical to Modern · Lesson 22 of 22

The Sensor Fusion Atlas

One map for the whole journey — every method, how they connect, and where they came from.

Prerequisites: the other 21 lessons (or curiosity). This is the index you return to.
24
Concepts mapped
64
Years of history
3
Interactive views

Tab 0: How to Read This Map

Twenty-one lessons is a lot of territory. You started by asking why two flawed sensors beat one good one, and you ended staring at transformers that fuse camera, LiDAR, and radar into a bird's-eye-view of a city street. Somewhere in between you derived the Kalman gain by hand, watched particles swarm a non-Gaussian belief, and learned why an information filter can be torn apart across a robot swarm and stitched back together with nothing but addition. This page is the single sheet of paper that holds all of it — the map you pin above your desk so the whole field fits in one glance.

It is, deliberately, lighter on long derivations than the lessons it points to. The value here is not new math; it is structure. Three things, in three views:

The one throughline that holds the whole field together

If you remember a single sentence from twenty-two lessons, make it this one:

Fusion is optimally combining information sources. Every method in this series — the humble inverse-variance blend, the Kalman gain, a particle reweighting, a factor-graph optimization, a transformer's cross-attention — is one answer to the same question: given several noisy, partial, possibly conflicting views of the world, how do I form the single best estimate, and how sure should I be of it? The methods differ in what they assume (linear or not, Gaussian or not, one robot or many) and in how they represent uncertainty (a covariance, an information matrix, a cloud of samples, a graph of constraints). But the question never changes. The whole atlas is variations on combining information.

And underneath the throughline is the discovery from Lesson 1 that reframes everything: different sensors fail in different places. A camera goes blind in the dark exactly where LiDAR shines; GPS dies in the tunnel exactly where the IMU keeps counting. Fusion is worth the trouble because the failure conditions don't overlap. Every architecture in this map is, at bottom, an arrangement for letting healthy information vote down the parts that have gone dark.

The four families — the spine of the map

The twenty-four concepts in the graph fall into four families, and they are not arbitrary buckets — they are four stages of generality. Each family answers a harder version of the same question than the one before it.

Family 1 — Combine measurements. The seed. You have two numbers (or two estimates) of the same quantity, each with a known uncertainty, taken at the same instant. How do you merge them? Answer: weight each by its precision — inverse-variance weighting — and the fused estimate is sharper than either input. This is Lessons 1–3: why fuse, where to fuse (the architecture / JDL question), and the two-measurement blend that every later method secretly contains. Master this and the Kalman gain is just this idea wearing a coat.

Family 2 — Recursive Bayesian filters. Now the world moves and measurements arrive over time. You can't re-fuse everything at every step. The Bayes filter is the recursive engine: predict the state forward with a motion model, then update it with each new measurement — carrying a belief, not a history. Everything else in this family is a way to make that engine tractable: the Kalman filter (linear-Gaussian, optimal, closed-form), the EKF and UKF (when the world is non-linear), the information filter (the same belief flipped inside-out into an additive form), the particle filter (when the belief isn't Gaussian at all), and covariance intersection (when you don't know how your estimates are correlated). This is Lessons 4–11 — the classical heart of the field.

Family 3 — Multi-sensor systems. A real robot doesn't run one filter on one clean signal. It must decide which measurement came from which object (data association — gating, GNN, JPDA, MHT), it must turn raw inertial readings into motion (IMU / inertial navigation, with its inexorable drift), it must bound that drift against an absolute reference (INS/GNSS coupling via the error-state KF), and before any of it works the sensors must be calibrated and time-synchronized — the unglamorous prerequisite that silently breaks everything when it's wrong. This is Lessons 12–15: the engineering that turns a filter into a system.

Family 4 — Perception fusion & the learned frontier. The modern era fuses rich, high-rate perception — whole images, whole point clouds — with inertial data. Visual-inertial odometry (MSCKF, VINS, IMU preintegration) and LiDAR-inertial / factor-graph systems (LIO-SAM, FAST-LIO2, iSAM2) reframe fusion as optimizing a graph of constraints rather than running a single recursive filter. Then perception goes end-to-end: multimodal driving fusion (camera + LiDAR + radar, early/mid/late) and transformer / BEV fusion (BEVFusion, TransFusion, BEVFormer) learn the fusion rule from data, and learned / differentiable filters (KalmanNet, BackpropKF) fold the classical recursion inside a neural network. This is Lessons 16–20.

Running alongside all four families is a fifth thread that isn't a family so much as a discipline: debugging and consistency (Lesson 21 — NEES/NIS, innovation whiteness, robust costs). It connects to every filter because every filter can lie to you about how sure it is, and a fusion estimate that is overconfident is more dangerous than one that admits it's lost. In the graph it's drawn as a diagnostic spine touching the methods it audits.

The family tree, in one block

Here is the whole lineage compressed — read top to bottom and you are reading the order the field was built in, and roughly the order the lessons teach it:

Family 1 · Combine measurements
Why fuse (L1) → Architecture / JDL (L2) → Inverse-variance blend of two measurements (L3). The seed: trust each source by its precision.
↓ add time & a motion model
Family 2 · Recursive Bayesian filters
Bayes filter (L4) is the trunk → Kalman (L5) → {complementary/AHRS (L6), EKF (L7), UKF (L8), information filter (L9)} → particle filter (L10), covariance intersection (L11). Carry a belief; predict then update.
↓ add many sensors, drift, real hardware
Family 3 · Multi-sensor systems
Data association (L12) · IMU / inertial nav (L13) · INS/GNSS error-state KF (L14) · calibration & time-sync (L15). The engineering glue.
↓ add rich perception, graphs, learning
Family 4 · Perception fusion & learned
Visual-inertial odometry (L16) · LiDAR-inertial / factor graphs (L17) · multimodal driving fusion (L18) · transformer/BEV fusion (L19) · learned differentiable filters (L20).
↻ audited throughout by
Diagnostics spine · Debugging & consistency (L21)
NEES / NIS, innovation whiteness, robust costs. Is your filter as sure as it claims to be? Touches every method above.

Three questions that route you through the whole map

When you face a real fusion problem, the map collapses into three questions — the same three from Lesson 1, now powerful enough to pick a method:

  1. Can one sensor produce the quantity alone? If no (depth from stereo, a 3D pose from a 2D image stream), you are forced into cooperative fusion — the quantity exists only jointly, and the methods live in Family 4. If yes, keep going.
  2. Do the sensors fail in the same conditions or different ones? Different conditions → complementary fusion (GPS+IMU; Families 1–2, the Kalman/complementary line). Same conditions, and a sensor might lie → competitive fusion (vote/cross-check; data association and consistency checks live here).
  3. Is it linear and Gaussian? If yes, the Kalman filter is optimal — stop. If non-linear, climb to the EKF or UKF. If non-Gaussian (multi-modal beliefs, kidnapped robots), reach for the particle filter. If you're fusing estimates whose correlation you can't track, use covariance intersection. If you have a window of past states to re-optimize, leave filtering for factor graphs.

Those three questions are the legend for everything below. The graph in Tab 1 is just these answers, drawn.

How to use this page. Treat Tab 1 as your hub: when a method feels isolated, open its node, read its one-line essence, and follow its edges to the ideas it grew from and the ideas that grew from it. Treat Tab 2 as the story: when you want to know why a method exists, find the problem of its era. Treat Tab 3 as the door out: every link goes to the paper or book where the idea was born.
What single question does every method in this series — from inverse-variance weighting to transformer BEV fusion — ultimately answer?

"What I cannot create, I do not understand — and what I cannot understand, I cannot teach."

Tab 1: Connections — The Concept Graph

This is the centerpiece. Every method in the series is a node, arranged in layers from foundations at the top to the learned frontier at the bottom. Solid lines are prerequisites ("learn this first"); dashed lines mean one idea evolved into or generalizes another. The teal diagnostic spine on the right audits the filters it touches.

Hover a node to light up its connections. Click a node to open its detail card below — its essence, a link to its lesson, and its key reference. On a phone, tap a node in the list that appears beneath the map.

The Sensor Fusion concept graph — click a node
hover to highlight · click to open a concept
Select a node

Click any concept in the graph above to read its one-line essence, jump straight to its lesson, and find the paper it came from. Start anywhere — the Bayes filter is the trunk that most of the classical methods branch from.

Every node, as a list

The same twenty-four concepts, grouped by layer. Tap any one to open its detail card above (and scroll up to the graph).

Reading the trunk. Notice how much of the graph hangs off one node: the Bayes filter. The Kalman filter is the Bayes filter assuming everything is linear and Gaussian; the EKF and UKF are the Kalman filter coping with non-linearity; the information filter is the same belief flipped into additive form; the particle filter is the Bayes filter when the belief refuses to be Gaussian. Four of the five classical methods are specializations of one recursion. That is the deepest structural fact in the field, and the graph makes it visible at a glance.
In the graph, an edge from the information filter to covariance intersection is drawn dashed, while the edge from the Bayes filter to the Kalman filter is solid. What does that difference encode?

Tab 2: Timeline — Sixty-Four Years of Fusion

The methods didn't appear at random — each answered a problem of its moment. The timeline below runs from Kálmán's 1960 paper to Barfoot's 2024 second edition, color-coded by era: classical (the linear-Gaussian foundations and tracking), robotics / probabilistic (particle filters, SLAM, inertial-visual systems, factor graphs), and deep / learned (differentiable filters and end-to-end perception fusion).

Click a milestone marker to read what it was, why it mattered, and which lesson it connects to. On a phone the timeline becomes a vertical scroll.

The history of sensor fusion — click a milestone
● classical ● robotics ● deep / learned click a marker
Select a milestone

Click any marker on the timeline to read its story. Start at 1960 — the year Rudolf Kálmán published the recursive estimator that the next sixty years would extend, linearize, sample, and eventually learn.

The three eras, in one breath. The classical era (1960–1990s) built the linear-Gaussian core and the tracking machinery for a few clean targets. The robotics era (1993–2018) embraced uncertainty that wasn't Gaussian (particle filters), tied estimation to a moving robot in an unknown world (SLAM, VIO, LIO), and reframed it as optimizing a graph (iSAM, factor graphs, preintegration). The deep era (2016–) makes the fusion rule itself learnable — either by folding a filter inside a network (KalmanNet, BackpropKF) or by letting a transformer learn how to fuse pixels, points, and radar into a bird's-eye view. Each era didn't discard the last; it wrapped it.
IMU preintegration (Forster, 2015) sits in the robotics era. What problem of its moment did it solve?

Tab 3: References — The Curated Library

The full bibliography for the series, in two groups: the textbooks and courses you'd keep on the shelf, and the seminal papers behind each lesson. Where a paper has an open arXiv copy or a stable DOI, it's linked.

Textbooks & Courses

  1. Thrun, Burgard & Fox. Probabilistic Robotics. MIT Press, 2005. The canonical text for Bayes/Kalman/particle filters, SLAM, and data association — the spine of Lessons 4–12.
  2. Barfoot. State Estimation for Robotics (2nd ed.). Cambridge University Press, 2024. Matrix-Lie-group estimation, batch vs recursive, factor graphs. The modern companion to Thrun. publisher
  3. Bar-Shalom, Li & Kirubarajan. Estimation with Applications to Tracking and Navigation. Wiley, 2001. The tracking bible — data association (PDA/JPDA), NEES/NIS consistency, IMM. Backs Lessons 12 and 21.
  4. Groves. Principles of GNSS, Inertial, and Multisensor Integrated Navigation Systems (2nd ed.). Artech House, 2013. The INS/GNSS reference — loose/tight/deep coupling and the error-state KF. Backs Lessons 13–14.
  5. Simon. Optimal State Estimation: Kalman, H-Infinity, and Nonlinear Approaches. Wiley, 2006. Clear derivations of KF/EKF/UKF and the information filter. Backs Lessons 5–9.
  6. Maybeck. Stochastic Models, Estimation, and Control (Vol. 1). Academic Press, 1979. The classic first-principles treatment of the Kalman filter and stochastic estimation.
  7. Särkkä. Bayesian Filtering and Smoothing. Cambridge University Press, 2013. Unifies KF/EKF/UKF/particle filters as Bayesian inference; covers smoothing. free PDF
  8. Dellaert & Kaess. Factor Graphs for Robot Perception. Foundations and Trends in Robotics, 2017. The factor-graph / smoothing-and-mapping (GTSAM) viewpoint behind Lesson 17. PDF
  9. Titterton & Weston. Strapdown Inertial Navigation Technology (2nd ed.). IET / AIAA, 2004. The strapdown-INS mechanization reference — how raw IMU becomes attitude and position. Backs Lesson 13.

Seminal Papers

  1. Kalman, R. E. "A New Approach to Linear Filtering and Prediction Problems." Journal of Basic Engineering, 1960. The paper that started it all (Lesson 5). DOI
  2. Julier & Uhlmann. "A New Extension of the Kalman Filter to Nonlinear Systems." SPIE AeroSense, 1997. The unscented transform / UKF (Lesson 8). DOI
  3. Julier & Uhlmann. "A Non-divergent Estimation Algorithm in the Presence of Unknown Correlations." American Control Conf. (ACC), 1997. Covariance Intersection (Lesson 11). DOI
  4. Gordon, Salmond & Smith. "Novel Approach to Nonlinear/Non-Gaussian Bayesian State Estimation." IEE Proc. F, 1993. The SIR particle filter (Lesson 10). DOI
  5. Bar-Shalom & Tse. "Tracking in a Cluttered Environment with Probabilistic Data Association." Automatica, 1975. The PDA filter (Lesson 12). DOI
  6. Reid, D. B. "An Algorithm for Tracking Multiple Targets." IEEE Trans. Automatic Control, 1979. Multiple Hypothesis Tracking, MHT (Lesson 12). DOI
  7. Madgwick, Harrison & Vaidyanathan. "Estimation of IMU and MARG Orientation Using a Gradient Descent Algorithm." IEEE ICORR, 2011. The Madgwick AHRS filter (Lesson 6). DOI
  8. Mahony, Hamel & Pflimlin. "Nonlinear Complementary Filters on the Special Orthogonal Group." IEEE Trans. Automatic Control, 2008. The Mahony attitude filter (Lesson 6). DOI
  9. Mourikis & Roumeliotis. "A Multi-State Constraint Kalman Filter for Vision-Aided Inertial Navigation." IEEE ICRA, 2007. The MSCKF (Lesson 16). DOI
  10. Qin, Li & Shen. "VINS-Mono: A Robust and Versatile Monocular Visual-Inertial State Estimator." IEEE Trans. Robotics, 2018. VINS-Mono (Lesson 16). arXiv:1708.03852
  11. Forster, Carlone, Dellaert & Scaramuzza. "On-Manifold Preintegration for Real-Time Visual-Inertial Odometry." IEEE Trans. Robotics, 2017. IMU preintegration (Lesson 16). arXiv:1512.02363
  12. Zhang & Singh. "LOAM: Lidar Odometry and Mapping in Real-time." Robotics: Science and Systems (RSS), 2014. LOAM (Lesson 17). DOI
  13. Shan, Englot, Meyers, Wang, Ratti & Rus. "LIO-SAM: Tightly-coupled Lidar Inertial Odometry via Smoothing and Mapping." IEEE/RSJ IROS, 2020. LIO-SAM (Lesson 17). arXiv:2007.00258
  14. Xu, Cai, He, Lin & Zhang. "FAST-LIO2: Fast Direct LiDAR-Inertial Odometry." IEEE Trans. Robotics, 2022. FAST-LIO2 (Lesson 17). arXiv:2107.06829
  15. Kaess, Johannsson, Roberts, Ila, Leonard & Dellaert. "iSAM2: Incremental Smoothing and Mapping Using the Bayes Tree." Int. J. Robotics Research, 2012. iSAM2 (Lesson 17). DOI
  16. Furgale, Rehder & Siegwart. "Unified Temporal and Spatial Calibration for Multi-Sensor Systems." IEEE/RSJ IROS, 2013. Kalibr — joint extrinsics + time offset (Lesson 15). DOI
  17. Vora, Lang, Helou & Beijbom. "PointPainting: Sequential Fusion for 3D Object Detection." CVPR, 2020. Point-level camera→LiDAR fusion (Lesson 18). arXiv:1911.10150
  18. Philion & Fidler. "Lift, Splat, Shoot: Encoding Images from Arbitrary Camera Rigs by Implicitly Unprojecting to 3D." ECCV, 2020. LSS — the camera-to-BEV lift (Lesson 19). arXiv:2008.05711
  19. Liu, Tang, Amini, Yang, Mao, Rus & Han. "BEVFusion: Multi-Task Multi-Sensor Fusion with Unified Bird's-Eye View Representation." ICRA, 2023. BEVFusion (MIT) (Lesson 19). arXiv:2205.13542
  20. Liang, Xie, Yu, Xia, Lin et al. "BEVFusion: A Simple and Robust LiDAR-Camera Fusion Framework." NeurIPS, 2022. BEVFusion (PKU) (Lesson 19). arXiv:2205.13790
  21. Bai, Hu, Zhu, Huang, Chen, Fu & Tai. "TransFusion: Robust LiDAR-Camera Fusion for 3D Object Detection with Transformers." CVPR, 2022. TransFusion (Lesson 19). arXiv:2203.11496
  22. Li, Wang, Li, Xie, Sima, Lu, Qiao & Dai. "BEVFormer: Learning Bird's-Eye-View Representation from Multi-Camera Images via Spatiotemporal Transformers." ECCV, 2022. BEVFormer (Lesson 19). arXiv:2203.17270
  23. Revach, Shlezinger, Ni, Escoriza, van Sloun & Eldar. "KalmanNet: Neural Network Aided Kalman Filtering for Partially Known Dynamics." IEEE Trans. Signal Processing, 2022. KalmanNet (Lesson 20). arXiv:2107.10043
  24. Haarnoja, Ajay, Levine & Abbeel. "Backprop KF: Learning Discriminative Deterministic State Estimators." NeurIPS, 2016. BackpropKF (Lesson 20). arXiv:1605.07148
  25. Sünderhauf & Protzel. "Switchable Constraints for Robust Pose Graph SLAM." IEEE/RSJ IROS, 2012. Robust back-end against outlier loop closures (Lesson 21). DOI
  26. Agarwal, Tipaldi, Spinello, Stachniss & Burgard. "Robust Map Optimization Using Dynamic Covariance Scaling (DCS)." IEEE ICRA, 2013. Dynamic Covariance Scaling (Lesson 21). DOI
  27. Solà, J. "Quaternion Kinematics for the Error-State Kalman Filter." Technical report, 2017. The error-state KF reference (Lesson 14). arXiv:1711.02508
The series, end to end. You began at "why fuse at all" and arrived at transformers that fuse a city block into a bird's-eye view. The methods multiplied; the question stayed the same. Combine your information sources optimally, and never trust an estimate more than it deserves. That is the whole atlas.