Back Matter

Glossary

The definitions this book relies on, written down once. In the chapters, a term’s first use links here with a dotted underline; here, each entry points back at the chapters that treat it. When a term has a precise meaning in this book, this page is where that meaning lives.

65 terms, first-use linked across 29 chapters

A

Area

The Lanelet2 primitive for an enclosed polygonal surface. Areas carry their own attributes and subtypes, so a parking lot, a container zone, and a non-drivable island are all Areas. Not every polygon that changes driving behavior is an Area: Autoware’s detection area is a regulatory element instead.

Covered in

Lanelet2 Primitives and Regulatory ElementsPart V · Ch 19

ASAM

The Association for Standardization of Automation and Measuring Systems, the standards body that maintains OpenDRIVE and the related OpenX family of automotive data formats.

Autoware

An open-source autonomous driving software stack built on ROS 2, covering localization, perception, planning, and control. This Handbook treats Autoware as the reference consumer of the maps it teaches you to build: when a map element has to earn its keep, the question is what Autoware does with it.

B

Behavioral contract

The agreement between a map and a planner under which encoded rules are treated as law. Geometry tells the robot where the road is; the regulatory layer tells it what it is allowed to do on it. The planner does not read signs, so the contract has no small print: what the map says is what the robot does.

C

Centerline

The reference line down the middle of a lane that planners follow and resample. In Lanelet2 the centerline is not stored geometry: it is computed from the left and right boundaries, which is why boundary errors show up as a shifted driving line rather than a broken one.

Cloth simulation filtering

A widely used ground-removal algorithm for point clouds: invert the cloud so the ground is the highest surface, drape a simulated cloth over it, and let it settle. The cloth’s shape is the ground; everything below the cloth is noise, everything above it is an obstacle candidate.

Covariance

The matrix attached to an estimate (a pose, a velocity) that quantifies how uncertain it is, and in which directions. Localization and fusion filters publish covariances alongside every estimate, and diagnostics watch them: a growing covariance is a pose going soft before it goes wrong.

D

Datum

The complete reference frame a coordinate is meaningful only in relation to: a specific ellipsoid, an origin, and an axis orientation. Two maps in different datums can differ by tens of meters while each being perfectly correct by its own rules. WGS84 is the datum of GNSS and of most modern mapping.

Dead reckoning

Estimating position by integrating measured motion forward from the last known-good fix instead of measuring position directly. IMU and wheel odometry dead-reckon between LiDAR fixes; drift is the method’s built-in tax, paid back every time an absolute fix arrives.

Detection area

An Autoware-specific construct (an extension regulatory element, not a Lanelet2 Area) paired with a stop line, that names a region the vehicle must check for crossing traffic or pedestrians before proceeding through an unsignalized intersection.

Digital twin

A scene-scale model of a real place, layered so that live state and sensor feeds can be projected onto it for operations and monitoring. Where an HD map is a contract for a robot, a digital twin is a dashboard for an operator; the two share capture data and coordinates.

E

Ellipsoid

The smooth oblate spheroid used as the reference surface for geodetic coordinates. The Earth is lumpy; the ellipsoid is the agreed smooth stand-in, and every lat/lon and ellipsoidal height is measured against one particular ellipsoid, most often WGS84’s.

ENU

East-North-Up: the axis convention for local Cartesian frames in ground robotics. X points East (eastings), Y points North (northings), Z points up. The mirror convention NED swaps to North-East-Down for aerial vehicles.

EPSG code

The standardized registry identifier for a coordinate reference system. An EPSG code names one exact system: projection, datum, units, axis order. Verifying that the map’s EPSG code and the planner’s match is the cheapest coordinate check in existence and the one most often skipped.

G

Geodetic coordinates

The coordinate triple (φ, λ, h) that locates a point relative to a reference ellipsoid: geodetic latitude, longitude, and ellipsoidal height. GNSS receivers report exactly this triple on WGS84. Ellipsoidal height is not elevation above sea level; bridging the two needs a geoid model.

Georeferencing

Anchoring a capture to global coordinates, typically via ground control points and a projection, so that the map sits where the world sits. A capture that skips or fumbles georeferencing is internally consistent and globally adrift: the floating-map problem.

Covered in

Scan ProcessingPart III · Ch 10 Grounding the Path: Tracing Reality in the PointcloudPart V · Ch 20

GNSS

Global navigation satellite system, the umbrella term for any satellite constellation providing positioning, navigation, and timing. GPS is one GNSS (the American one); GLONASS, Galileo, and BeiDou are the other global systems. Modern receivers track several at once, which is why the precise habit is to say GNSS.

GPS

The United States’ Global Positioning System: the first and best-known satellite navigation constellation. In current practice GPS means either the American system specifically or consumer-grade single-constellation positioning; this Handbook says GNSS for the general capability.

Grid and ground coordinates

The two flavors of a projected coordinate system. Grid coordinates are the raw projected values; ground coordinates are scaled so that distances in the data match distances measured on the asphalt. The two differ because flattening the Earth stretches it, and confusing them scales every layer subtly against every other.

Ground control point

A physically marked point whose coordinates are known from survey, used to anchor a capture to the Earth and to check it afterwards. Ground control is what turns a pretty point cloud into a georeferenced one; at Harbor Yard it is one survey nail plus six fixed landmarks.

Ground truth

Reality as it physically is, independent of any map, sensor, or estimate. Ground truth does not care what your file says: it is the reference that validation compares against, and the reason a map that looks right in the viewer can still be wrong.

Covered in

What Is an HD Map?Part I · Ch 1 From Physical World to Autonomous MachinePart I · Ch 4 The Same Object in Multiple RepresentationsPart IV · Ch 13

H

Harbor Yard

The running example site of this Handbook: a forty-acre logistics depot with a warehouse, loading docks, a container zone, and a one-way loop road, served by an autonomous cargo tug at 5 mph. Every concept in the book is exercised there first.

HD map

A high-definition map: a lane-level, machine-readable representation of a place, precise to centimeters and carrying semantics and rules, used by a robot for localization, planning, and prediction. The defining property is not resolution but trust: the planner treats it as the environment.

I

ICP

Iterative closest point, the foundational scan and cloud registration algorithm: repeatedly pair each point to its nearest neighbor in the other cloud, solve for the rigid transform that best aligns the pairs, and iterate. Slow to converge from a bad initial guess, and the reason registration pipelines start coarse.

IMU

Inertial measurement unit: a sensor measuring linear acceleration and angular velocity at hundreds of hertz. An IMU never knows where it is, only how it is moving, which makes it the perfect partner for a slower absolute fix: speed between LiDAR matches, corrected by them.

L

Lanelet

The atomic primitive of a Lanelet2 map: a drivable lane segment defined by its left and right linestring boundaries, a direction of travel, and attributes such as subtype and speed limit. Lanelets connect to each other through declared successor relationships, forming the routing graph.

Covered in

Lanelet2 Primitives and Regulatory ElementsPart V · Ch 19

Lanelet2

An open-source, lane-level map format for autonomous driving, built on OpenStreetMap primitives: points, linestrings, lanelets, areas, and regulatory elements. It carries geometry, topology, and rules in one file, which is why it answers both of the planner’s questions: where can I go, and what am I allowed to do along the way.

Covered in

What Is an HD Map?Part I · Ch 1 The Many Maps of the Same PlacePart I · Ch 3 The Vector Survey: Choosing Autoware Lanelet2Part V · Ch 18 Lanelet2 Primitives and Regulatory ElementsPart V · Ch 19 The HD Mapping Workflow and ValidationPart VI · Ch 24 How Autonomous Vehicles Use MapsPart VII · Ch 32

LiDAR

Light detection and ranging: a sensor that fires laser pulses and measures their time of flight, producing a dense cloud of 3D points. LiDAR is the geometric backbone of HD mapping: it drives capture, localization scan matching, and perception’s object detection alike.

Linestring

The polyline primitive of Lanelet2 and OpenStreetMap: an ordered sequence of points. Linestrings are pure geometry until tags make them mean something: the same linestring can be a lane boundary, a stop line, or a curb, depending on its attributes.

Local site frame

A Cartesian coordinate frame defined once for a single site: an origin at a convenient point (often a survey marker) and axes aligned to the site or to ENU. Coordinates stay small, distances are plain meters, and there is no projection distortion; the trade is that the frame means nothing outside the site.

Localization

The robot’s continuous answer to “where am I”: estimating the vehicle’s pose by matching live sensing (typically LiDAR scans) against the pre-built pointcloud map, fused with fast inertial and odometry sensors. Localization accuracy bounds everything downstream, because a planner can only be as right as the pose it plans from.

Loop closure

The SLAM mechanism that recognizes a previously visited place and adds a constraint tying the current estimate back to the earlier one, collapsing accumulated drift across the whole loop. Capture routes are designed around loop closure: crossing your own path is not wasted driving, it is insurance.

M

Map matching

Aligning a live measurement against a map to determine where the vehicle is: the umbrella term that includes LiDAR scan matching against a pointcloud map and map-relative positioning generally. In navigation parlance it can also mean snapping a GNSS trace to a road network; in this book the LiDAR form is the one that matters.

MGRS

The Military Grid Reference System: a tiling scheme layered on UTM that divides the world into zones, squares, and addressable grid cells. In robotics deployments MGRS is used as the tiling scheme for pointcloud map tiles, which is where its one-map-one-zone rule comes from.

N

NDS

The production delivery format for automotive maps (and its successor NDS.Live), maintained by an automotive consortium: a compiled, tiled, versioned database built for shipping maps to millions of vehicles rather than authoring them. Author in Lanelet2, deliver in NDS, is a common split.

NDT

Normal distribution transform scan matching: the algorithm Autoware’s pose estimator uses to align a live LiDAR scan against the pointcloud map. It models the map as a grid of local Gaussian distributions and searches for the transform that best aligns the scan’s points with them, producing a pose plus a covariance.

NED

North-East-Down: the axis convention used mostly by aerial vehicles, where Z pointing down matches the direction cameras and rangefinders face in flight. NED and ENU describe the same world; converting between them is a flip and a rotation.

O

Occupancy grid

A rolling grid of cells around the vehicle, each marked free, occupied, or unknown. Together with the lanelet map it answers “where can I go”: the lanelets say where driving is permitted, the occupancy grid says where it is physically clear, and the unknown cells are the blind spots.

OpenDRIVE

The ASAM road description format for simulation: a road network as centerline geometry with offsets, signals, and objects, authored to drive traffic simulators rather than robots. Many projects keep one source of truth and export both a Lanelet2 map for the vehicle and an OpenDRIVE export for the simulator.

OpenStreetMap

The collaborative, openly licensed map of the world. For HD mapping it is two things: a source of prior knowledge about a site before you scan it, and the data model (nodes, ways, tags) that Lanelet2 builds its primitives on.

OpenUSD

Open Universal Scene Description: a layered, composable scene format from which rich 3D content is assembled. In this book’s stack it is the structure behind the custom site assets layered onto the map and the natural backbone of digital-twin builds.

P

Perception

The robot’s answer to “what is around me”: detecting objects in the current frame, tracking them across frames, and predicting where they go next. Perception leans on the map twice: the pointcloud map is subtracted as known background, and the vector map constrains the forecasts.

Covered in

How Robots Understand PlacesPart I · Ch 2

Planning

The subsystem that reads the map’s connectivity and rules to answer “where can I go” with an actual path: a route over the routing graph, refined into a trajectory that respects every rule and obstacle along it. Planning is where the map proves its worth, and where its gaps surface as behavior.

Covered in

How Robots Understand PlacesPart I · Ch 2 From Physical World to Autonomous MachinePart I · Ch 4

Point cloud

A set of points in 3D space, typically millions of them, each one a measured coordinate. The point cloud is the geometric ground truth of a capture: precise, dense, and meaningless until semantics and rules are layered on top of it.

Pose

The combined position and orientation of one coordinate frame expressed in another: translation plus roll, pitch, yaw. The vehicle’s pose (vehicle frame within map frame) is the quantity localization estimates and everything downstream consumes.

Prediction

The perception stage that forecasts where each tracked object is likely to go, as a set of candidate paths with confidences. The map anchors the forecast by supplying the candidates themselves: prediction asks the vector map which lanelets the object could occupy next. If the correct future is not on the menu, it cannot be predicted.

Projection

The mathematical operation that converts angular coordinates on the ellipsoid into planar Cartesian ones measured in meters, so that planners can compute distances with plain subtraction. Every projection stretches the Earth somewhere; choosing, documenting, and using one projection everywhere is the discipline.

R

Refline

The reference line a planner builds along its route and aligns constraints against: where a stop line bites, where a speed zone begins. When a regulatory element cannot produce a refline, planners fall back to a conservative default, which is one way a map bug becomes a mystery stop.

Registration

Aligning multiple scans or captures into a single consistent coordinate frame, pairwise (ICP, NDT) or globally (pose-graph optimization). Registration is where capture quality is won or lost: a registration error is invisible in any single scan and present in all of them.

Regulatory element

The Lanelet2 construct that encodes a rule: a first-class object tying a traffic sign, signal, or stop line to the lanelets it governs. Lane geometry says where you may drive; regulatory elements say what you must do there. An orphaned one, whose referenced geometry has been deleted, is the classic cause of a ghost stop.

Covered in

Lanelet2 Primitives and Regulatory ElementsPart V · Ch 19 Behavioral Contracts: Regulatory Elements and Site PolicyPart V · Ch 22 The HD Mapping Workflow and ValidationPart VI · Ch 24 How Autonomous Vehicles Use MapsPart VII · Ch 32 When the Map Breaks the RobotPart VII · Ch 36

ROS 2

The Robot Operating System 2: the publish-subscribe middleware and tooling that most modern robot software, Autoware included, is built on. The four questions of this book map onto ROS 2 nodes and topics you already know if you come from that world.

Routing graph

The directed graph compiled from lanelet successor relationships, over which the planner computes routes. The routing graph is why geometric adjacency is not a connection: if the topology declares no edge, the route does not exist, however obviously the lanes touch.

RTK

Real-time kinematic GNSS: differential positioning carried to its centimeter conclusion by resolving carrier-phase measurements against a base station of known position. RTK is how a capture drive pins its trajectory (and often the map) to surveyed coordinates in real time.

S

Scan matching

Estimating pose by finding the transform that best aligns a live scan with the map (localization) or with another scan (registration). NDT and ICP are the two algorithms this book follows; both need an initial guess and both degrade in feature-poor environments.

Semantics

The layer of meaning attached to captured geometry: which cluster of points is a building, which traced line is a lane boundary, which polygon is a crosswalk. Semantics are what turn a 3D model a robot could render into a map it can reason about.

SLAM

Simultaneous localization and mapping: estimating the sensor’s trajectory and the map at the same time, each making the other harder. Modern HD mapping narrows SLAM’s scope with GNSS, RTK, and loop closure, but the chicken-and-egg problem at its core is why capture is not just driving around with a scanner.

Covered in

Simultaneous Localization and MappingPart III · Ch 11 Simultaneous Localization and MappingPart III · Ch 11 The HD Mapping Workflow and ValidationPart VI · Ch 24

T

Topology

The connectivity structure of a map: which lanelets feed which, where routes exist and where they dead-end. Topology lives in declared relationships, not in drawn geometry, and planners trust it over geometry: two touching lanelets with no topological edge are not connected.

Total station

The surveyor’s electronic theodolite: measures angles and distances to a target to produce coordinates at millimeter relative accuracy over site scale. For HD mapping it is the instrument of record for ground control and for anything that must be right beyond argument.

Trajectory

The planner’s output: a path with a time parameter attached, giving position and velocity along it for the controller to track. The map’s constraints (speed limits, stop lines, clearance) all land here, transformed from map data into the shape of the drive.

Triangulation

Determining an unknown position from measurements to two or more points of known position: the surveyor’s classic method, and the conceptual ancestor of both GNSS positioning and landmark-based localization.

Covered in

Map-Based LocalizationPart III · Ch 12

U

UTM

Universal Transverse Mercator: the standard projected coordinate system dividing the world into 6-degree zones, each with its own meters-based grid. UTM is the default world frame for HD maps, including Harbor Yard’s (zone 10N); its one-map-one-zone rule exists because each zone is its own projection.

Covered in

Coordinates, Frames, and Reference SystemsPart II · Ch 5 Geographic and Projected Coordinate SystemsPart II · Ch 6 Geographic and Projected Coordinate SystemsPart II · Ch 6 The Global Alignment: Reality Meets ReferencePart IV · Ch 17 The Global Alignment: Reality Meets ReferencePart IV · Ch 17 The HD Mapping Workflow and ValidationPart VI · Ch 24

V

Vector map

The semantic road-network layer of the map, in this book the Lanelet2 map, as distinct from the pointcloud map it was traced over. Localization consumes the pointcloud map; planning and prediction consume the vector map; the two must agree because they are the same place.

Voxel

A volumetric pixel: one cell of a 3D grid imposed on a point cloud. Voxelization trades exact points for tractable structure, powering ground removal, occupancy grids, and the voxel-based detectors in perception.

W

WGS84

The World Geodetic System 1984: the reference frame and ellipsoid GNSS positions are reported in, and the effective common language of global coordinates. Most projected map frames, UTM included, are anchored back to WGS84.

Covered in

Coordinates, Frames, and Reference SystemsPart II · Ch 5 Geographic and Projected Coordinate SystemsPart II · Ch 6 The Global Alignment: Reality Meets ReferencePart IV · Ch 17

Wheel odometry

Estimating vehicle motion by counting wheel rotations at high rate. Like the IMU it is fast and relative, and like the IMU it drifts (wheels slip), which is why both are fused with an absolute fix rather than trusted alone.