Skip to main content

Asseti

Asseti wanted its platform to find and measure parking lines in drone imagery, a job its people were doing by hand. We placed one of our L6 engineers in the Asseti team for six months to build the first version: a computer vision model, trained on YOLO and refined with OpenCV, that finds each painted line, grades its condition and counts the metres of paint, with very high accuracy, all shown on the map in real time.

About Asseti

Asseti monitors the condition of buildings and sites. Licensed drone operators capture imagery of a property, and Asseti's platform, AssetOS, turns it into 2D and 3D models, finds and ranks defects, and builds a register of the assets on the site. Owners and facilities managers of commercial, industrial and retail property use it to see the state of their sites and plan maintenance.

What Asseti needed

Car parks are a large part of many of the sites Asseti looks after, and their painted lines wear out. To report on them, people had to trace each line by hand on the drone imagery, grade its condition and work out how many metres of paint the car park held. On a large site that is slow, repetitive work, and two people tracing the same car park will not always agree.

Asseti wanted the platform to do this itself: find every parking line in the imagery, say how worn it is and count the metres of painted line, so that a repaint could be planned and priced from the results. It needed an experienced engineer who could work with imagery and maps as well as code, and who could take a first version from idea to something that worked on real sites.

How we helped

We placed one engineer in the Asseti team for six months from 1 November 2025, at L6 on our grading, the level above lead engineer. Our engineer worked within Asseti's team on the first implementation of parking line detection, from the first experiments to a model that performed well on real sites. The engineer remained our employee throughout, so we stayed responsible for employment and support while Asseti directed the work.

The work had four parts: preparing the drone imagery, training a computer vision model to find the lines, cleaning up and measuring what the model found, and showing the results on a map in Asseti's web app. In short, we trained a computer vision model (YOLO) to find each painted line, then used OpenCV and algorithms we wrote ourselves to clean, grade and measure them.

Working with GIS imagery

A drone survey produces hundreds of overlapping photographs. These are stitched into a single orthomosaic, an aerial image corrected so that its scale is the same across the whole site, and georeferenced, so that every pixel has a known position on the earth. This is GIS imagery: a picture that is also a map.

That matters for every step that follows. A line found in the image can be placed exactly on the site's map, and its length can be measured in real metres rather than pixels. A single site image is far too large for a model to read at once, so it is cut into overlapping tiles. Each tile is processed on its own, and the results are transposed back into the site's map coordinates and joined together, so that a line crossing the edge of two tiles comes back as one line.

Training the model

YOLO, short for You Only Look Once, is a family of computer vision models widely used across the industry to find objects in images. Older approaches scanned an image many times, looking at one region after another. YOLO looks at the whole image in a single pass and predicts where every object is and what it is at the same time, which makes it both fast and accurate. It comes in sizes from nano to extra large. Larger models have more layers and parameters, so they can learn finer detail at the cost of more computing power. We used the large model, because thin, faded paint on tarmac is exactly the kind of fine detail a smaller model misses.

Models like this are not built from nothing. The industry practice is to start from a model already trained on millions of everyday images, which has learnt to recognise edges, shapes and textures, and then train it further on examples from the job at hand. That second stage depends on labelled data: images in which people have marked every parking line by hand. The labelled images are split into a training set, which the model learns from, and a validation set it never sees during training, which shows whether it has really learnt to find lines or has only memorised the examples.

To help the model cope with real sites, the training images are varied on purpose. They are rotated, flipped, brightened, darkened and partly hidden, so that the model learns to find a line whether it runs north to south, sits in shadow or is half covered by a parked car. The model is then trained over many rounds, checked against the validation set after each one, and the best version is kept.

Cleaning up the results

On its own, a model's output is close but not clean. A line may come back in several pieces, with a ragged edge, or with a stray mark next to it that is not paint. To turn that into results Asseti could rely on, we built a post-processing pipeline in several steps, using OpenCV, the standard open-source library for image processing, together with algorithms we wrote for the job.

The pipeline removes small marks that cannot be parking lines, smooths each line's edges and thins it down to a single centre line. It then joins the pieces of a broken line into the line it belongs to, using the direction and spacing of the lines around it, because parking bays follow regular patterns. Finally it straightens each line and fits it back to its true position on the map. Each step makes the output cleaner, and together they took the model's results to a very high level of accuracy.

Each clean line is then graded. Comparing the paint still visible along a line with the line as it should be gives a measure of its condition, so a line that is crisp and complete can be told apart from one that has faded or worn through. Because the imagery is georeferenced, the length of every line is measured in real metres, and the lengths are added up to give the total metres of painted line on a site.

Showing it on the map

The results only help if people can see them. The lines, their grades and their lengths are layered onto the site's imagery in Asseti's web app, so that a user can look at a car park, see every line coloured by its condition and select any line to see its details.

A single site can hold a very large image and thousands of lines, so the app has to be extremely efficient to stay responsive. The imagery and the line layers are loaded in pieces, only for the part of the map on screen and at the level of detail needed, so that the map, its layers and all their information appear in real time as the user pans and zooms.

The result

By the end of the engagement Asseti had a working model that finds, grades and measures parking lines with very high accuracy. Work that needed a person to trace and grade every line by hand is now done by the platform, which saves a great deal of time on each site and gives the same answer every time the same imagery is checked.

It also gave Asseti a base to build on. With line detection, grading and measurement working, the same approach can be applied to other markings and surfaces that Asseti's clients need to look after.

Timeline

  1. Nov 2025

    The engagement starts, with one of our L6 engineers joining Asseti's team to work on parking line detection

  2. Nov 2025 to Apr 2026

    The large YOLO model trained on labelled drone imagery, with a multi-step OpenCV pipeline to clean, grade and measure each line, and the results layered onto the map in the web app

  3. Apr 2026

    Six months on, Asseti has a working model that finds, grades and measures parking lines