Pipeline localisation using camera

[MadsJJ]

Motivation

We implement a pipeline localisation approach using the camera(s) to find a precise location of the pipeline starting point. Camera(s) is picked in favour of sonar because the sonar has significantly more noise. Because we are underwater, the assumed range is max 10 meters

Description

Using the image from the camera(s), determine whether it contains the pipeline. If the pipeline is displayed, locate the starting point. The starting point of the pipe can be assumed to be endpoint closest to the drone or origo.
Planning to use YOLOv8 since we get confidence scores. It also gives us a segmentation mask.

Task suggestion

• Create a YOLO training script vortex-deep-learning-pipelines#30
• Create training data from simulator for pipeline segmentation using front facing camera vortex-auv#618
• Train a YOLOv8 model to segment the pipeline based on a front facing camera image vortex-deep-learning-pipelines#22
• Locate pipeline starting point using segmentation mask #19
  • Can use pipe point closest to drone
  • If we use a single monocular camera, we wont have depth. A 3D position can for example be estimated by assuming a flat ground plane and using the DVL altitude estimate

Deliverables

• A method to determine wether the pipeline is present in the image
• A method to create a segmentation mask on the pipeline the image
• A method to locate the starting point of the segmented pipeline

References

• Real world data of pipe from tacc: https://studntnu.sharepoint.com/:u:/r/sites/o365_Vortex2021/Shared%20Documents/Software%20-%20Perception/Data/rosbags/2024/tacc2/[tacc2_0.db3](https://studntnu.sharepoint.com/:u:/r/sites/o365_Vortex2021/Shared%20Documents/Software%20-%20Perception/Data/rosbags/2024/tacc2/tacc2_0.db3?csf=1&web=1&e=8rVqxm)?csf=1&web=1&e=8rVqxm

Contacts

@MadsJJ