Practical machine learning to estimate traffic flow in San Juan, Puerto Rico

santurcesoftware.com

61 points·boredumb··6 comments

In short, the author used yolo5 to count cars from webcam streams locations along highways in San Juan. I think it would have been easier to do background subtraction and count vehicles using standard vision techniques (e.g., threshold, watershed, count blobs), given that they had access to a fixed view security feed.

I think the author used Yolov5 to detect cars and their own method for counting them. Counting requires tracking object instances, which Yolo doesn't do.

He rant it every 60 seconds, so I don't think he was too concerned about tracking instances. And it doesn't really matter for purposes of classifying traffic levels.

Counting the different vehicles to boot strap it and have it start saving training images into the categories so that a new model could be trained against those images, and not rely on object detection to create new training data or power that app.

For the night images, I would imagine just number of red pixels would give a roughly accurate count which could be converted into "high" or "low"

Despite having worked in technology based jobs for 20 years and been involved with computers for almost 30, it still boggles my mind that we've reached a point where we can:

- connect to publicly available video streams - over public internet - run existing computer vision libraries - to identify real objects correctly

and all for free!

It really is astounding.

Not sure if some of this comes from working in old school FinTech where code is old, monkey patched etc or just my age. I will say that this is the kind of HN that keeps me inspired that tech will just keep getting easier and easier to use which will in turn lead to cooler and cooler projects.

Reply on news.ycombinator.com