Open a LiDAR scan of an ordinary street and you get millions of dots. No labels. A lamp post is a tall, thin cluster. A manhole cover is a faint disc sitting flush with the road. For years a technician spotted each one by eye and drew it by hand.
Machine learning now takes a first pass at that job. It’s also one of the few AI use cases you can check against a drawing, which makes it a good place to see what the technology does well and where it stops.
A scan holds positions, not meaning
A scanner records where things are: x, y, z, often colour and reflectivity too. It never records what they are. “Manhole” is an idea that a person, or a model, has to bring to the data. Everything below is about supplying that idea automatically.
The seven objects AI finds today
Detection tools work on a short list of street furniture and utility objects. VisionSpot, an AI tool built for surveying, lists these as available now:
- streetlights
- road signs
- utility poles
- manhole covers
- storm drains
- trash cans
- fire hydrants
Its developers plan to add trees, benches, electrical cabinets, traffic lights and general infrastructure poles. Planned isn’t shipped, so check the current list before you promise anything to a client.
Why point objects come first
A lamp post is a discrete thing with a clear location. A kerb line is a continuous feature whose start, end and corners take judgement. Detection works best on the first kind, which is why every object on the list ends up as a single symbol on a plan. The tool’s own page calls the job point-object extraction, and the wording is accurate.
The output is a list of detections with positions. From there a drafter converts the confirmed ones into blocks in AutoCAD or Civil 3D, using a library for the topographic format the client wants.
How a model learns what a hydrant looks like
Nobody writes a rule for it. The model is shown labelled examples, scans where each feature has already been marked, and it picks up the geometric characteristics that define the class. Then it applies that to point clouds it has never seen. Training data is the expensive part, because someone has to mark the examples. After that, new files get processed at the speed of the computer rather than the speed of a person scrolling.
That matters because simple rules fail in predictable ways. “Tall, thin, vertical” describes a pole. It also describes a tree trunk and a downpipe.
The reality check
Detection proposes. A person decides. In VisionSpot’s workflow you upload a cloud, run detection and select which detected objects go forward, so review is part of the process rather than an afterthought.
Expect misses where a parked car covers a drain or an object is partly hidden. Expect odd local variants that the training examples never covered.
And treat time-saving figures as claims to test. The team says its own engineers save from 15% of project time, depending on project size. Your result depends on your data and on how much review you need.
Try it on your own file
The best test is a real one. VisionSpot reads LAS, LAZ and E57 files, from 500 MB up to 100 GB, and the developers offer to run a sample you send, configured to your drafting standards, free of charge. Details are at gis-jot.com/visionspot-ai.
Seven object types is a modest list. Each one is repetitive tracing that nobody will miss.
