Laborelec, a subsidiary of Engie, is a research and expertise center in electrical power technology.
In order to optimise the purchases related to their maintenance operations, Laborelec was looking for a solution to count, classify, and geolocate the street lamps in specific urban areas. Laborelec knew it was possible to infer these numbers from the video footage coming from the cameras installed on their service cars. However, extracting this data from the videos manually was too expensive to be done manually.
To help Laborelec, Sagacify developed a machine learning model to automatically detect, classify, and geolocate street lamps, based on video streams from in-car videos.
Once the algorithm extracted the required data, the procurement department was able to order exactly the right quantity of spare parts, which improved the maintenance process and storage costs.
To maximise the impact of the available labelled data and reduce the dataset to a maximum, Sagacify set an active learning system up, which ensured that the data proposed for labellisation would have a maximal impact on the model performances.
Thanks to this approach designed to minimize the cost of labellisation, the final model used only 700 labelled images, which were labelled in a tool specifically deployed for the client, in a total of less than 4 hours.
The final model accurately detects 99% of streetlights with a classification accuracy of more than 90%.
Using the video streams from service cars proved to be an efficient method of inferring the geolocation, quantity, and type of street lamps.