Locating lampposts

Municipality of Amsterdam

The purpose of the system is to update the registration of lampposts. This registration is used for asset management (including maintenance) of public lighting.

Last change on 27th of November 2024, at 17:04 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: Field not filled in.
• Status: In use

General information

Theme

• Space and Infrastructure
• Traffic

Begin date

Field not filled in.

Contact information

[e-mail]

Responsible use

Goal and impact

The purpose of the system is to update the registration of lampposts. This registration is used for asset management (including maintenance) of public lighting. The system detects lampposts in 3D point clouds. It then calculates properties of the lamp posts, such as the exact location of the bottom and top of the lamp post, and the angle it makes with the ground.

Considerations

Field not filled in.

Human intervention

Each found lamp post (and its associated fit) is checked by a human.

The information about the found lamp posts that are easy to match with the existing registration is automatically included in this registration.

The found lamp posts that are not so easy to match with the existing registration are checked again in more detail by experts, before the information on these posts is used.

Level of human control: human-in-command

Risk management

The system is low risk because the data is not very sensitive and the outcomes are checked by people before we use them.

Large scale data processing

• Risk description: Large scale data processing (segmentation of everything that can be seen from the road, sense of being part of a machine, automated system)
• Frequency: Low
• Description of risk mitigation: Communication. We explain that we do not use everything and that everything is controlled by humans.
• Mitigation status: Completed
• Likelihood: Low
• Scale: Low
• Severity: Low

Location data

• Risk description: Location data (knowing where everything is might be helpful for people with malicious intentions)
• Frequency: Low
• Mitigation status: Not started
• Likelihood: Low
• Scale: Low
• Severity: Low

Non-discrimination: We expect the model to perform best in areas similar to those the model is trained on (in this case, same shapes lampposts and appearance of the environment). We want to try to keep the impact of (the quality of) the model the same for different groups. Therefore, Amsterdam Oost was chosen as the location for training the model, as this district has a diverse streetscape and population.

Operations

Data

Name: 3D point clouds

Dataset description

The input data are the 3D point clouds made with a LiDAR scanner in 2020. The points have x, y, z coordinates, a colour and intensity.

The points are first semantically segmented and then clustered into individual lamp posts.

To train the model, a subset of the data was used, from 50 pieces in Amsterdam East. This subset was annotated. This was partly done automatically and then corrected by a human. Initial tests were carried out in Weesp and the Oosterpark neighbourhood in Amsterdam Oost. The final operational model uses the dataset of all of Amsterdam.

For the automatic annotation, we also use AHN4 and BGT data to determine, for example, the land and location of buildings.

• Licence: right of use within the municipality
• Operating
• Personal data:No personal data

Source:

Street LiDAR | Cyclomedia

Technical design

Description of the system architecture:

The first step is the semantic segmentation of point clouds. Here, we classify points into regions based on their meaning. We determine whether something is part of a lamppost or not. For this, we use RandLA-Net (QingyongHu/RandLA-Net (github.com)).

The next step is to cluster all points that belong to lampposts. We use connected-component labelling for this purpose. We remove noise and eventually arrive at pieces of point clouds that belong to an individual lamppost.

For each cluster found, we use prinicipal component analysis to create a fit for each lamp post. From this fit, we can easily derive the properties of the pole.

The code for these last two steps can be found on our github: Amsterdam-AI-Team/Urban_PointCloud_Analysis

Performance

• Performance first step in data processing (semantic segmentation with RandLA-Net): the intersection-over-union for the lamppost class is 82.
• Performance second step (identifying individual lampposts): a significant proportion (about a third) of the lampposts found are in reality not lampposts (false positives). These are picked out by humans (see heading 'human monitoring').

Performance third step (fit per lamp post): in 91 per cent of cases the fit was correct and in the remaining 9 per cent was corrected by a human.