Autonomous vehicle perception operates on two timescales simultaneously. On one timescale, real-time perception models process sensor data at 10–30 Hz to detect and track dynamic objects in the environment, other vehicles, pedestrians, cyclists, obstacles. On another timescale, a static map provides the structural context that real-time perception depends on where roads are, how lanes are connected, where intersections are, what traffic control devices exist and where.
High-definition maps are the artifact that captures this static context at the precision that AV systems require. HD map annotation, the process of labeling map elements with the semantic meaning, geometric precision, and connectivity information that AV systems need is a distinct annotation discipline from the object detection and tracking annotation that most AV discussions focus on. It produces a different type of ground truth, requires different expertise, and supports a different set of AV model capabilities.
What HD Maps Contain That Standard Navigation Maps Don’t
Consumer navigation maps Google Maps, HERE Maps, OpenStreetMap at standard resolution provide road geometry at the precision needed for turn-by-turn navigation directions. AV systems need substantially more: lane-level geometry, semantic attributes of each lane, precise positions of traffic control devices, and the topological connectivity between lane segments that tells the AV planning system which paths through an intersection are legal and how they connect.
HD map layers that AV systems require:
Lane geometry: The precise centerline of each lane segment, accurate to 20–50 centimeters in horizontal position. This precision allows the AV system to localize itself within a specific lane (not just a road) and to plan paths that respect lane boundaries.
Lane attributes: For each lane segment, semantic attributes that affect driving behavior lane type (driving lane, turning lane, bicycle lane, parking lane), lane width, speed limit, direction of travel, allowed maneuvers (straight, left turn, right turn), and lane change legality (whether the lane boundary is solid/no-change or dashed/change permitted).
Road boundary geometry: The physical edges of the road surface curbs, barriers, road edges where there is no curb at the same precision as lane geometry. Road boundary geometry defines the navigable space beyond which the AV system should not plan.
Intersection topology: The connectivity graph that describes how lane segments connect through intersections which entry lanes connect to which exit lanes, the geometry of each connecting path, and any conditional movement restrictions (protected left turn only when signal indicates).
Traffic control devices: The precise position of traffic signals (including which specific approach lanes they govern and their physical mounting geometry), stop signs, yield signs, speed limit signs, and other regulatory signs. Position precision matters because the AV system uses map-registered signal positions to associate signal detections from the camera with the specific intersection movement they govern.
Lane markings: The type, position, and extent of painted road markings, lane boundary lines, stop lines, crosswalk markings, turn arrows, speed limit pavement markings.
Why HD Map Annotation Requires Different Expertise
HD map annotation is not image annotation applied to aerial or LiDAR survey data. It is a GIS-adjacent discipline that requires understanding of road topology, traffic engineering conventions, and the specific semantic attributes that AV planning systems need.
An annotator who correctly places bounding boxes around vehicles in camera images may not have the knowledge to correctly label an intersection’s lane connectivity graph which entry lanes connect to which exit lanes, accounting for the lane channel assignments that are sometimes implicit in the physical geometry rather than explicitly marked. Getting this connectivity wrong produces a map that the AV’s planning system uses to generate incorrect or illegal maneuver sequences.
Annotators for HD map work need:
- Road engineering literacy: Understanding of lane numbering conventions, turn lane assignment logic, intersection control design, and the standard conventions for road markings in the target geography.
- GIS tool proficiency: HD map annotation uses specialized GIS and vector annotation tools (JOSM, ArcGIS, proprietary AV map annotation platforms) rather than the image annotation platforms used for camera annotation.
- Geographic context: Road conventions differ substantially across countries and even regions within countries. Intersection geometry, sign design, marking conventions, and signal placement all vary. Annotators need familiarity with the specific geographic context of the roads being annotated.
- Connectivity reasoning: The intersection topology layer requires reasoning about the connectivity of a complex three-dimensional structure often from aerial imagery or LiDAR survey data that shows the physical geometry but not explicitly the connectivity structure. Annotators need to infer connectivity from the physical geometry using their understanding of intersection design.
The Annotation Process for HD Maps
HD map annotation typically begins with survey data, a dedicated survey pass of the target road network using a high-precision mobile mapping system that captures LiDAR point clouds with centimeter-level positioning accuracy, camera imagery with GPS synchronization, and inertial measurement data. This survey data provides the geometric foundation that annotators label with semantic meaning.
The annotation process layers semantic content onto the geometric survey data:
Lane geometry extraction: Annotators trace lane centerlines through the point cloud and imagery, placing the center point of each lane at the correct lateral position across the range of the survey. Tools that support snapping to detected lane marking positions in the LiDAR data accelerate this process for well-marked roads; roads with degraded or absent markings require more manual positioning judgment.
Attribute assignment: For each mapped lane segment, annotators assign the semantic attributes lane type, speed limit, allowed maneuvers, change permissions drawing on both the physical evidence in the survey data (visible turn arrow markings, lane width, sign positions) and knowledge of the traffic engineering conventions of the target geography.
Connectivity graph construction: For each intersection, annotators explicitly connect each entry lane to each exit lane through the intersection, specifying the path geometry of each connection. This connectivity graph is the most labor-intensive part of complex intersection annotation and the most consequential for planning correctness.
Traffic control device placement: Annotators place point features at the precise physical positions of each traffic signal, stop sign, and regulatory sign, linking each device to the lanes and movements it governs.
Map Change Detection and Update Workflows
HD maps degrade over time as the physical road environment changes, new road construction, temporary lane configurations for events or maintenance, sign replacement, signal timing changes. A map that was accurate at survey time may be partially inaccurate six months later in areas that have seen infrastructure changes.
Map change detection for AV systems uses the AV fleet’s own sensor data: cameras and LiDAR on operational vehicles continuously observe the road environment and compare it against the current map. Detected discrepancies, a sign that was in the map is no longer present, a lane marking that wasn’t in the map has been added, a construction zone that created a new temporary lane configuration generates change candidates that are queued for human review and map update.
This continuous update workflow requires its own annotation infrastructure: change candidate classification (is this a real change or a detection error?), change characterization (what specifically changed?), and map update annotation (updating the specific map elements affected by the change). The latency between a road change occurring and the updated map being validated and deployed to the AV fleet determines how long the AV system operates with a map that doesn’t match the current road environment.
How Map Quality Affects Real-Time Perception
HD maps improve real-time AV perception performance in several specific ways, each dependent on map accuracy:
Prior for detection: A map that specifies where lanes are provides a prior for where vehicles and cyclists should be detected. Detection models that incorporate map context correctly localize their detections relative to known lane geometry. Map inaccuracy produces a biased prior that degrades detection localization.
Motion prediction context: Trajectory prediction models that incorporate lane connectivity predict lane-adherent trajectories more accurately than models without map context. Map connectivity errors produce incorrect trajectory predictions for lane-following behavior the most common driving behavior.
Localization: AV localization systems align real-time sensor observations against the HD map to estimate the vehicle’s precise position. Localization accuracy is bounded by map accuracy; a map that is geometrically incorrect introduces systematic localization error.
Final Thought
HD map annotation is the AV training data discipline that is least visible in AV Perception Model Training Data discussions but most foundational for the full system’s operation. The lane geometry, connectivity topology, traffic control placement, and semantic attributes that HD maps provide give real-time perception models the structural context that transforms their point-in-time sensor observations into understanding of a permanent, navigable road network.
The quality of that map its geometric accuracy, its semantic completeness, its connectivity correctness bounds the quality of every perception capability that depends on it. Programs that invest in HD map annotation expertise, with annotators who understand road engineering conventions and connectivity reasoning, build the static foundation that dynamic perception models rely on to understand where they are and where they can go.