Inspection of oil and gas infrastructure
- Data processing
- a week → 1 hour
- the report is ready at base
Task: inspect routes and facilities from a helicopter without processing terabytes of video by hand after the flight.
How it works
- 01The problemInspection by eye and review of the recordings after the flight: it takes a long time, requires processing a large volume of data, and still is not accurate enough
- 02The solutionComputer vision in real time on board. Detection of objects and anomalies during the flight, computed by edge AI on the device itself — power and space rule out the cloud and an external server
- 03The resultInspection without an operator on board: vision on an NVIDIA Jetson chip runs the survey in real time and produces the final report as soon as the helicopter is back at base
A YOLO cascade for people and vehicles, GroundingDINO / SAM for equipment and anomalies. Offline mode with local weights, tracking with event deduplication.
Particulars and difficulties
The cloud is not available
Power and space on board rule out an external server, so everything is computed on the device
There is no link in the field
Offline mode with local weights: the inspection does not depend on a channel that will not be there
One object, many frames
Tracking with event deduplication, so an object crossing the frame is one event
Two classes of target
A YOLO cascade for people and vehicles, GroundingDINO and SAM for equipment and anomalies
Results
- a week → 1 h
- to process the data
- 3 → 1
- people in the process
- 0
- operators on board
The report exists on landing
It is produced during the flight rather than a week after it
Inspection without an operator
Vision on the chip runs the survey itself, and the crew shrank from three to one
Stack
NVIDIA Jetson (L4T) · PyTorch · ONNX · YOLO · GroundingDINO · SAM