Object Recognition Public Space

Municipality of Delft

Automatic recognition and counting of objects or features in public spaces.

Last change on 25th of February 2026, at 11:33 (CET) | Publication Standard 1.0

• Publication category: Other algorithms
• Impact assessment: DPIA, AIIA
• Status: In use

General information

Theme

Space and Infrastructure

Begin date

Field not filled in.

Contact information

[e-mail]

Responsible use

Goal and impact

The municipality of Delft applies image recognition to make public space measurable. This is done by capturing street images and analysing these images using image recognition techniques. This application provides information about the physical condition of public space and is deployed in various forms as part of smarter and data-driven management and maintenance of public space. The municipality of Delft uses this technique to determine the quality level based on assessment criteria (e.g. number of litter lumps detected). The technology of image recognition in combination with image measuring sticks makes it possible to efficiently gain insight into the state of maintenance of public spaces.

Process: 1. Image acquisition

Image acquisition is carried out by employees or hired persons, using recording equipment (such as GoPro's or 360° cameras). These are attached to vehicles, bicycles, or carried by persons on foot. Recording is only done from public roads or other locations where image taking is allowed. If the routes are planned, in cooperation with a municipal client, this client may choose to inform residents in advance via, for example, the municipal website, newsletter or neighbourhood newspaper. The cameras primarily target relevant objects in public spaces, such as roads, trees and street furniture. Nevertheless, potentially identifiable elements may unintentionally come into view; however, these are not the purpose of the recording.

The outcome of the analysis is a dataset consisting of numerical scores, count values or area-specific classifications (e.g. by street section or grids). After collection, the images are transferred to a secure storage environment as soon as possible, and no later than 24 hours, or uploaded directly via an encrypted connection to the secure on-premise server environment.

Anything that can trace back to individuals is blurred. Blurring is the process of obscuring or blurring certain parts of an image or video so that it is no longer traceable to individuals.

2. Destruction of raw images on recording equipment

After successful transfer of images to UrbanVue's secure server environment, the raw images on the storage media used (such as SD cards) are immediately destroyed.

The deployment of cameras is not intended to replace current practices, but to support and enhance them.

Considerations

Images of public space containing personal data (if any) are blurred within 24 hours and these anonymised images are automatically deleted after 2 weeks. The cameras are not deployed structurally.

The data collected in this way can be used efficiently for the maintenance and cleaning of outdoor spaces. The quality of the outdoor space is improved, which in turn leads to a more liveable environment.

The supplier has investigated the availability of less intrusive alternatives (such as aerial photos, residents' reports, manual inspections), but concludes that they do not offer an equivalent level of detail, timeliness or operational scale.

- Aerial or satellite images offer insufficient resolution to identify litter, defective street furniture or maintenance levels.

- Reports from residents are subjective, fragmented and not systematic.

- Manual inspections require a lot of physical presence on the street, which - although not digital - can also have privacy implications (structural observation without safeguards), and are also costly, slow and not scalable.

The deployment of cameras is not meant to replace current working methods, but to support and strengthen them. The professionalism of staff and the involvement of residents through reports remain valuable - the question is how we use this scarce capacity as effectively as possible.

The deployment of camera images with automated analysis offers the opportunity to: Target skill deployment: routine inventory is supported by image recognition techniques, allowing staff to focus on assessment, prioritisation and complex situations where their expertise is really needed. Use scarce capacity more effectively: instead of going through the entire area manually, employees can work in a targeted way based on up-to-date data. Following up resident reports better: with an up-to-date and comprehensive picture, we can prioritise and deal with reports faster. Control costs: structurally hiring extra staff for area-wide inspections is costly and unrealistic in the current labour market. Automated support offers a scalable alternative. The cameras are not intended for surveillance of individuals, but for imaging the physical state of public spaces.

Human intervention

For verification purposes, a manual sample is taken periodically on a representative selection of images. This involves checking whether automatic anonymisation has been carried out correctly. If this check reveals that certain elements have not been anonymised or are incomplete, these are corrected before further processing takes place.

There is no automated decision-making by using the image recognition facility.

Risk management

- Only images of public spaces (or areas where it is allowed) are processed; no courtyards or private areas;

- Cameras are deployed incidentally per project, not structurally or permanently;

- Recording is done only in areas relevant to the assignment (e.g. selected neighbourhoods or measurement routes), not randomly everywhere;

- Personal data (faces, license plates) are automatically blurred within 24 hours;

- Blurred images are automatically deleted after 2 weeks;

- UrbanVue does not perform profiling, tracking or behavioural analysis on individuals;

- Final products contain only anonymised or aggregated information.

Legal basis

AVG Article 6(1)(e): public interest

Impact assessment

• Data Protection Impact Assessment (DPIA)
• AI Impact Assessment (AIIA)

Operations

Data

Objects or features in public spaces (such as litter, weeds, crookedness or broken elements) are recognised and counted.

Any personal data (faces, license plates) these are automatically blurred within 24 hours.

Technical design

The vendor uses on-premise servers and algorithms that systematically detect faces and vehicle license plates and other potentially identifiable objects in the images. These elements are immediately pixelated, in a manner that is irreversible (i.e. not reconstructible by reasonable means). The anonymised images are then analysed using object detection and classification algorithms. These algorithms recognise and count objects or features in public space (such as rubbish, weeds, crookedness or broken elements). The outcome of this analysis is a dataset consisting of numerical scores, count values or area-specific classifications (e.g. by street section or grids).

External provider

UrbanVue B.V.