Galadrim implemented an on‑premises computer vision solution to strengthen safety on a barrel‑filling line at Castrol’s Péronne plant.
Castrol, a subsidiary of the BP group and a specialist in industrial lubricants, operates a factory in Péronne. As part of strengthening its safety standards, Castrol launched a pilot project at these facilities.
The objective was to implement a computer vision system to monitor a packaging line (barrel filling) and detect any human intrusion in real time.
This project addressed two main challenges:
Operator safety:
ensuring the protection of teams working near the automated line.
Regulatory compliance:
following a quality audit that highlighted a safety gap, Castrol needed to deploy an effective alerting solution to meet the Group’s requirements.
To meet Castrol’s needs, Galadrim developed a complete solution combining hardware and software.
Training a computer vision model
We trained a YOLO (You Only Look Once) object‑detection model, specialised in detecting people and their postures. To ensure reliability, the model was trained with a dataset created specifically for the project:
Analysis of 4 hours of video from the production line provided by the client.
Manual annotation of 500 images to teach the model to identify intrusions under real factory conditions.
Hardware deployment
The hardware solution includes an NVIDIA GPU to run the detection model, as well as a control screen so operators can view alerts instantly.
Implementing this solution meant overcoming several technical and environmental constraints specific to the factory.
No Internet connectivity
The packaging line had no Internet access. It was therefore essential to design a self‑contained system. We addressed this by using a dedicated NVIDIA GPU capable of running the computer vision model entirely locally, without relying on the cloud.
Physical installation
Camera positioning was critical to ensure complete visual coverage of the line. After a site visit, our teams advised Castrol’s staff on installing a custom support system, positioned above the line so the cameras could be fixed in the optimal location.
Hardware configuration
The project involved significant hardware engineering, notably the advanced configuration of NVIDIA cards to optimise real‑time detection.
