Ccv now has a state-of-art tracking algorithm | Hacker News Reader

Ccv now has a state-of-art tracking algorithm

libccv.org

82 pointsliuliu12 comments

Specifically, the state of the art algorithm mentioned is the Predator/TLD developed by Zdenek Kalal.

There is a C++ port of the original MatLab code (more than one actually - search look at the forks of the original OpenTLD project on github), but I much prefer C code myself - thanks, liuliu!

The tracking red square in this video reminded me that ever since I played GR:AW and watched Macross+, I dream of a human-driven car that would give more information to the driver in the form of a full-size windshield+HUD overlaying reality with outlines and visible/hidden information. This HUD would display tangible data whether they're seen or not (i.e pedestrians crossing behind a parked van or passengers passing in front of a bus at a bus stop, or at night, roadside outlines, etc...) and intangible ones like velocity vector, stop distance, and projected trajectory, with possible collision alerts.

Although BMW has started something like this, the UI they came up with looks like this disaster of a UI[2][3][4], whose crappiness makes it a pricy toy, not a real security accessory (yet).

[0]: http://img.ebw.gr/13536/ghost-recon-advanced-warfighter-scr....

[1]: http://www.play3-live.com/screens/ghost-recon-advanced-warfi...

[2]: http://www.wired.com/images_blogs/autopia/2009/10/bmw_night_...

[3]: http://static.ddmcdn.com/gif/in-dash-car-night-vision-system...

[4]: http://www.theautochannel.com/news/2007/02/06/036291.1-lg.jp...

That's actually pretty exciting, I love having an alternative to OpenCV (which, while extensive, has many faults)

Looks super interesting, I had need for something similar a while back so I thought cool, let's try this out. But I am unable to find any sample code, is there any available? Can somebody point me in the right direction if so. I have not worked with ccv before so feels like a bit high learning curve without any samples.

Would be really cool if there were a rough guide on how the motorcycle tracking was done so that I could know in which end to start.

there are some CLI sample code in ./bin/

That's fantastic!

When the box disappears does that mean the algo has completely lost track of the rider though?

If I am not mistaken then, yes, it does mean that. However, unlike many other trackers, when the object comes back in view, it will pick it up again - it learns how the object looks, and will continue to look for it.

It's kind of mind-blowing that there are no false-positives (that I saw) with the rapidly moving ground.

There were a few false positives in the demo video (occasionally catching the engine block from the side), but it extremely impressive.

That's a pretty easy track though - a high contrast shape that wasn't changing shape (only position in the frame) that much.

When it did change i.e. due to perspective and angle to the camera, the tracker lost it.

Planar tracking would allow tracking with perspective and size changes.

Very good work. I'm interested in the vision technology and have seen it progress rapidly in recent years. The tracking is pretty amazing.

Liu you're an animal. Nice work!

Reply on news.ycombinator.com