Ben Cantrick (mackys) wrote,
Ben Cantrick

  • Mood:
  • Music:

An honest to god all-nighter.

So, my consulting work is interesting. It's also very time-consuming. I've spent since 1pm this afternoon more or less continuously working on getting a big batch of Excel charts made and printed. (It's just after 4am now.)

Making the charts was bad enough. There were 60-odd of them to do, and they seemed to take 10-12 minutes each. You can do the math. But even printing the damn things has turned out to be an exquisite pain in the ass, too. Basically, the printer is streaking and blurring the right side of most pages. I think it's some kind of wax-transfer job and some roller inside is overheating. Thus the streaking. (Remind me again why it's so damn hard to get a computer job? Because all the computers and associated gear are all working so goddamn great, right?) I thought I could just queue up the print jobs and let it run. Instead I'm sitting here and my employer is paying $0.20/min so the printer can cool down between individual pages. I put $80 on a Kinkos payment card before I started, thinking that would be comfortably overkill. But it's going to be razor-close...

This stuff is all due at 7am, in a little less than three hours. (I'm having breakfast with the guy I'm working for.) By the time I'm done with breakfast I'll have been awake for more than 24 hours continuously. After THAT I get to go on-shift at Radio Shack from 11-9. (Special extended holiday hours. Yay... yay...) And after that I'll attempt to be coherent for abelits's birthday party. It's been a long-ass day, and it's going to be another one...

Since I'm bored and have nothing better to do than sit here and type at 4:30 in the morning, let me tell you a little about the work. I can't tell you exactly what it is, of course, but I think I can give you a rough idea.

Imagine, if you will, that we didn't have traffic lights. Instead, there was a traffic director at every intersection. The traffic director makes various motions with his hands to tell people in cars to stop, go, turn left, turn right, etc.

Now, suppose the cars become driven by a computer. But we have all these human traffic directors, that for complicated bureaucratic reasons, they don't want to get rid of - or even re-train. They want the computer driven cars to recognize the gestures the traffic directors make, and respond accordingly.

Sounding like a difficult problem in AI yet?

The approach taken was multi-pronged. First, they decided to put special gloves with built-in accelerometers on the hands of the directors. Second, a computerized vision system that will attempt to "see" the gestures, or at least something, will be on the car. And there's a third method in the mix too - voice recognition. Reliable speaker-independent voice recognition is hard. Throw in a noisy, highly variable environment like these directors work in and it's even harder.

If you're engineering minded at all, at this point alarm bells are probably going off in your head. "Big, complicated system. Inherently hard problems. Serious threats to the lives of the directors if the system should malfunction. Big, fat, hairy, ugly problem."

My end of it is, I think, probably the easiest at least in theory. My job is to take the acceleration data from the gloves and make a computer program that will try and recognize the patterns as the accelerometers swing, and match them up to gestures.

The best way I can figure to do this is to do a numerical double-integration on the accel data to get position data. Then look for patterns in the position data. It'd be nice if you could just look at the straight accelerometer data, but unfortunatly I don't think that's going to work. There are a lot of gestures in the set we need to recognize that involve very similiar accelerations. For example, the gesture for "turn right" involves raising your left arm parallel to the ground, then raising your right hand like you were saying the pledge of alliegance, and waving the right hand back and forth as if to say goodbye.

Seems pretty easy, right? Both hands get short periods of accel as they move into position, then the right hand accelerates back and forth while the left hand says pretty much still. Here's the problem though - there's another gesture (put your left hand on your nose and waggle your right index finger back and forth) that looks, in acceleration terms, essentially completely identical to the turn right gesture.

So the double integration has to happen, and then we're stuck with what's essentially a simple handwriting recognition problem in 3D. I'm hoping to fake it with some curve-fitting stuff, but even that's somewhat sketchy. I'm seeing weird stuff in the data - the accelerometers are maxing out during certain gestures, which means the integrated position data won't be accurate.

It's gonna be interesting, that's for sure...

Bloodhounds shall cruise. Rendevous. Break curfew, new trip ensues.
You never know, Bloodhounds just go. Pack backpack up for down the road.
No change of clothes but change for tolls. Pack of No-Doz, Rolos and Skoal.
Old Milwaukee, Jolt, black coffee, Sugar Daddy, Rand McNally.
A pack of jacks, those orange Tic-Tacs, Glen Miller tracks, roll of Kodak.
The phantom black stacked Pontiac. Hot rod dual quads and in the back...
Is the beat keeper - big ass speaker. Jumbo woofer, rattle tweeter.
Dial is turned and there it's stayin'. The next day where Stern is playin'.

You got your Jesus on the dashboard, but the devil's under my hood.
You're taking it down legal, I'm pullin' it up to no good.
God is your co-pilot - I let Satan ride shotgun.
You pay a toll to get to heaven, but on the road to hell there's none.

Get up you're asleep at the wheel!
Get up, you're asleep at the wheel!

  • Post a new comment


    default userpic

    Your reply will be screened

    Your IP address will be recorded 

    When you submit the form an invisible reCAPTCHA check will be performed.
    You must follow the Privacy Policy and Google Terms of use.