Facebook IconFacebook IconFacebook IconFacebook Icon

AI Powered Street lights Detection

Sagacify Logo


Laborelec, a subsidiary of Engie, offers services to the energy industry (electricity producers, grid operators, …), to communities, and to the public sector. A subset of these services includes the census of the number, type, and condition of the street lamps deployed in Belgium.

Until recently, this task was performed manually, and the quality and cost of the data gathered was found to be suboptimal. Indeed, this is a repetitive, long and tedious task, which makes it prone to errors. To solve this problem, Laborelec partnered with Sagacify to develop a machine learning model to automatically detect and classify street lamps on video streams from cameras fixed on cars.

There were several challenges for this project.

  • Deep learning algorithms are well known for their exceptional performances at objects detection and classification. But these algorithms also need a large set of labelled data to be trained accurately.
  • There was no labelled dataset available, and the labellisation work had to be kept to a minimum.
  • The distribution of the different types of street lamps is highly unbalanced which makes the classification task harder.
  • Depending on the background, some street lamps can be complicated to spot, even for humans.


To create a labelled dataset, Sagacify configured and deployed an efficient labelling environment for Laborelec. With this tool, Laborelec employees were able to label the videos efficiently.

In order to minimize the labellisation efforts, our model used transfer learning to minimise the number of labelled data needed to perform this task. We also used active learning to feed the labellisation tool with the images which would accelerate the training process as much as possible.

This methodology ensures that the labelling task was kept to a minimum for the best possible model performances.


With 761 labelled images, our model reached 88% of precision on the detection of street lamps on single images. As one would expect, this precision increases as the lamps are closer to the camera.

This precision proved to be more than enough to spot all the street lamps on the video streams since the lamps missed on one frame have a high probability of being detected on the next frames of the video.

Our model also reached more than 95% of correct classification of the lamps on single images.

Our client now only has to drive in the streets to have precise vision of the number, type and location of street lamps as they drive by.

Related cases