Point Cloud to CAD

Converting an aerial LiDAR point cloud into a CAD drawing takes five steps. First you check that your data supports the drawing you have been asked for. Then you classify the cloud, so the software knows which points are ground, which are buildings and which are trees.

From the ground points you build a terrain model. From the terrain model and the classified points you extract vectors: contours, spot heights, building outlines, road edges and everything else the drawing needs. Finally you organise those vectors onto named layers and export them as DXF.

Classification is the step that decides the quality of everything after it. Get it right and the rest is mostly configuration.

Search for “point cloud to CAD” and almost everything you find is about a different job. Most results cover terrestrial scanning: a tripod scanner or a handheld unit captures a building interior, and the goal is a 3D model of walls, floors, pipes and structure. That work is usually called scan to BIM, and it is measured in rooms.

Aerial LiDAR is a different problem with a different answer. Your scanner was on a drone or an aircraft looking down. Your survey area is measured in hectares, not rooms. Nobody wants a 3D model of a hillside.

What the client wants is a survey drawing: contour lines at a stated interval, spot heights, the outline of every building, the edge of the road, the watercourse, the trees, all correctly georeferenced and all on layers their own CAD template recognises. So the tools are different, the outputs are different, and advice written for scan to BIM will send you down the wrong path.

Classification assigns every point in the cloud a label describing what it hit. This is the foundation of the whole conversion, because every vector layer you extract later is derived from a class. Contours come from the ground points. Building outlines come from the building points. Tree crowns come from the high vegetation points.

If a wall is labelled vegetation, no amount of care at the export stage will put that building on your drawing. Modern classification is done by a machine learning model trained on aerial LiDAR rather than by hand-tuned filters.

Good software lets you fix these by selecting points and reassigning them, using a brush, a height threshold, a colour range or an intensity range, rather than by editing anything point by point. Budget time for this pass. It is usually short, and it is the difference between a drawing that gets accepted and one that comes back.

With the ground class correct, the terrain model is next. A Digital Terrain Model, or DTM, is a continuous surface representing bare earth, built by interpolating between ground points onto a regular grid. Two decisions matter.

Resolution. The grid cell size should follow your ground point density. Cells much smaller than the average spacing between ground returns do not add information, they just make a bigger file and a rougher surface that produces wobbly contours. Cells much larger than the spacing throw away detail you paid to collect.

What gets excluded. The DTM is built from ground points only. Buildings, vegetation and vehicles are excluded, which is exactly why classification had to come first. A surface that still includes roofs is a Digital Surface Model, a different product entirely, and contours drawn from it will run over the tops of buildings.

Everything downstream flows from this surface. Contour lines are traced from it. Spot heights are sampled from it. The elevation attached to a building footprint is read from it. A terrain model built on a misclassified ground surface produces a drawing where every elevation is quietly wrong, which is far more expensive than one where something is obviously missing.

This is where the point cloud becomes drawing objects. Each layer is extracted independently, so you request only what the deliverable actually needs.

All of these export as GeoJSON, Shapefile and DXF, so the same extraction serves a GIS handoff and a CAD handoff without reprocessing.

A DXF full of correct geometry on one layer called 0 is not a deliverable. The export is where the drawing becomes usable.

Layer organisation. Every feature type goes on its own named layer: major contours separate from minor, buildings separate from roads, each utility layer separate. If your client has a layer naming standard, apply it at export rather than asking them to remap twenty layers by hand.

2D or 3D. Contours and building outlines can be flat polylines at a stated elevation, or true 3D polylines carrying Z. A drafting workflow usually wants 2D with elevation as an attribute or a label. An engineering workflow modelling surfaces wants the Z. Decide before you export, because converting afterwards is tedious.

Coordinates and units. Confirm the exported drawing is in the coordinate system and the unit your client's template expects, then open the file and measure a known distance. This takes a minute and catches the error that otherwise gets discovered by somebody else, later, at your expense.

Format choice. DXF is the interchange format and opens everywhere, which is why it is the sensible default for a handoff. Shapefile suits a GIS recipient. GeoJSON suits anything web based. Send the format the recipient actually works in.

Open the drawing and work through this list. It takes a few minutes and it catches nearly everything.

• Do contour lines cross each other? They physically cannot. Crossings mean a problem in the terrain surface, usually a classification error.
• Are contours running across building roofs? Your surface is a DSM rather than a DTM, or buildings were not excluded.
• Do buildings sit at sensible ground elevations? A structure floating above or sunk below the terrain points at a classification or interpolation problem.
• Measure a known distance. This catches the unit and coordinate system errors.
• Are all the layers present and named correctly? Compare against what you agreed to deliver, not against what came out of the export.
• Is anything on the drawing that should not be? Vehicles, noise returns and temporary site objects are the usual suspects.
• Does the drawing sit in the right place? Overlay it on a known reference and confirm the origin.

Because the classification, the terrain model and the vectors are all produced in one pass, changing your contour interval or adding a layer does not mean starting again. New accounts include 50 hectares of free processing, which is enough to run a real site end to end and see the DXF before deciding anything.