Mapping intersections and roads is a task for the brave.

An OSM way is a line without width, so any lane model has to assume where that line sits on the pavement — by default, down the middle of the carriageway. Wherever the road is not symmetric about its own geometry, that assumption produces wrong output. On a segment where a turn lane is added, the pavement widens on one side only, but the centered line pushes every lane sideways. Forks, merges, and the short two-node stubs between a ramp and the mainline fail the same way.

placement states where the centerline runs: middle_of:2, left_of:1, right_of:3 pin it to a numbered lane or lane edge; transition marks a way whose line is not parallel to the lanes, as at a fork, merge, or lane-count change. OSMPIE adds two extensions. dist:2.5 gives the offset in meters — positive to the right in the direction of travel, negative to the left — and placement:forward / placement:backward set each direction separately. Two centerlines offset outward render a buffer zone without a false lanes:both_ways=1; a negative offset carries them past each other so left turns pass on the right.

Lane positions are derived from the centerline, so if the line is assumed centered when it is not, every lane, stop line, and conflict point downstream is placed on the wrong part of the road.

Most lane connectivity is never written down — it is inferred. lanes gives the count, turn:lanes gives each lane’s permitted movements as pipe-separated values read left to right in the direction of travel, and width:lanes gives the widths. Together they yield a lane-level picture of the junction: where the incoming and outgoing ways carry the same number of lanes, lanes are taken to continue one to one, and widths plus placement put those lanes where they actually sit on the ground.

The inference breaks down where the lane count changes across a junction, where several roads merge, or where markings oversimplify what drivers do. turn:lanes says which direction a lane turns; it does not say which lane of the target road that lane enters. The connectivity relation says it — from way, via node, to way — listing each source lane and its destination lanes, with parentheses for a destination reachable only by changing lanes. It supplements turn:lanes, never replaces it.

A lane-by-lane model of a road network is a graph of lanes, and without connectivity — inferred from counts and widths, or stated by relation — there are only disconnected lane fragments with no edges between them.

A conflict zone is the part of an intersection where nobody may stand. It is the area that has to be left clear, that a driver must not enter unless they can leave it, and that a pedestrian steps into at their own risk — because it is exactly where the paths of everyone crossing the intersection meet. Draw every maneuver a driver, rider, or pedestrian can make through the node, and the conflict zone is the shape holding all of their crossing points. Outside it, no two paths conflict.

junction:radius gives the radius, in meters, of the circle that this shape is inscribed in, centered on the intersection node itself. One number on the node covers all lanes.

With the radius set right, the circle lands where the intersection really begins: stop lines and other entry points sit on it. Set it too small and the circle no longer contains the crossing points — 4 m on an intersection of wide multilane roads leaves maneuvers conflicting outside it, and the description is simply wrong. Neighboring nodes within the radius pull the zone out of round, since they constrain the possible maneuvers. Where lanes need to enter at different depths, junction:radius:lanes overrides the node value lane by lane, with forward/backward and start/end suffixes selecting side and approach direction.

A lane-by-lane model of an intersection needs to know where each lane stops and where the shared crossing area starts, and the radius is what fixes that boundary.

The functional zone of an intersection is the area where objects fall under its influence or influence it in turn — parking, crossings, and stop lines are interpreted differently inside that zone than outside it. junction:cluster:radius gives the radius, in meters, of a circle that this zone fits into. Buffer each tagged node by its radius and union the results: overlapping circles merge into one polygon, and that polygon is one junction instead of a scattering of separate nodes. Where the tag is absent, 12 meters is assumed.

A relation or a string cluster key would also group nodes, but both demand manual work on every ordinary junction — referential integrity, uniqueness control, generated tags. A radius is a plain number in meters, geometric by nature, and it creates no new object type: the junction is derived on demand by buffer plus union. It also works in reverse. A small radius keeps a nearby node from merging — separating an uncontrolled crossing that sits close to a signalized intersection but is not part of it.

A junction is a set of nodes joined by ways within which traffic control provides no stopping of vehicles — one indivisible conflict zone — so without a defined cluster you cannot tell which nodes share one stop line, one signal object, or one set of intersection calculations.