Teaching computers to see
Propane tanks have a unique shape that makes them instantly recognisable—both to people and to computers. This unique shape, when combined with aerial photography, allows the latest computer models to automatically identify tanks over vast areas.
This technique is called object detection. It’s a part of the field of computer vision, in which computers examine images and videos to see details people might miss. And, in this case, to complete tasks quickly that would otherwise take a huge amount of time.
To apply this technique, you need several things. First is the required data. In this case, that’s aerial photographs. You then need to teach the computer model where propane tanks are located in those photographs using labels. Once this has happened, the real learning can begin.
The learning process for a computer is similar to that of a human. As a child, your parents may have pointed to a bicycle and told you what it was. When you have heard and seen this a number of times, you develop the ability to recognise bicycles yourself.
Similarly, when training a computer, you show it several examples of the object you want it to identify. When it incorrectly identifies an object, you give a ‘penalty’ helping the model to improve as time goes on. You then continue this process until there is no room left for improvement.
That doesn’t mean that an object detection model will be perfect, however. It may not recognise propane tanks obscured by other objects; those under a tree, for example. But, compared to a manual approach, it will quickly enable you to scan large areas in a short amount of time.
Our propane tank detection model
So, what does the end result look like.
In the proof-of-concept (POC) phase of our own propane tank detection project, developed in collaboration with RUD Zeeland, we used aerial photographs with a resolution of 5cm sourced from an environmental service.
After showing this entire area to a computer, our model then provided a list of coordinates that may contain propane tanks, along with pictures of the top 500 results that could be verified by a user.
Unsurprisingly, we found this process makes it a lot easier and far faster for an employee to ascertain where propane tanks are located. Especially when compared to searching the area in person or scouring Google maps by eye.
Our assessment of the POC, is that this application has significant potential, and we are now looking to develop it further alongside any interested parties.
Your next steps
At Royal HaskoningDHV Digital, we have a great deal of experience with object detection. We’ve previously used it extensively to identify wet cooling towers, applying false positives to dramatically improve the performance of our model. And we expect our propane tank model to show the same levels of improvement.
This solution will enable organisations to search for propane tanks in a fast and efficient way. In doing so, they can ensure compliance with the latest environmental laws, but also make sure that tanks are regularly maintained and that our environment is protected.
If you’d like to know more about this technique and what it could mean for you, get in touch.