Fitting two xy sequences of data

[Kahr]

Hi Alex,

Yes the angle map is the point-point-point vector angle when moving along the path of the contour points.

It is derived using IPAN99 methode.

I am not sure I understand this question.

"how are these "real world objects" transported in the active area of the camera?"

It is simply a grey scale image with background illumination.

The syringes are transported on a conveyor system and an image is grabbed on the fly.

[Kahr]

Hi Altenbach

That looks great. I will definitively use this approach for the correlation.

It is not given that the defect always is in the same location. It can appear anywhere on the contour.

[alexderjuengere]

---

@altenbach wrote:

---

---

Of course you could map the data into 2D arrays and do a 2D cross correlation directly. 😉

---

I was about to do that

thank you!

@Kahr wrote:

Hi Altenbach

That looks great. I will definitively use this approach for the correlation.

It is not given that the defect always is in the same location. It can appear anywhere on the contour.

 this appears to be the proper and necessary pre-processing step  before calculating

any similarity between template and sample.

"It is not given that the defect always is in the same location."

That's what I basically had in mind, but was limited by my german tongue

when I asked those cryptic question in Message 8

[Kahr]

Hi

I've had the time to analyze Your method Altenbach and I am actually back at my original problem - cross correlation is (of course) time consuming and slow.

The generation of the 2D arrays of 600x600 points brings me kind off back into the image domain 

On my i7 laptop it takes 350 ms. Using traditional edge based PM in VDM takes 50 ms.

So that was the reason for taking a 1D approach with less data points.

Kind regards

[altenbach]

---

@Kahr wrote:

On my i7 laptop it takes 350 ms. Using traditional edge based PM in VDM takes 50 ms. 

So that was the reason for taking a 1D approach with less data points.

---

That's less than an order of magnitude in speed and I intentionally used 2D array sizes that are way too large, with lots of zeroes on all sides.

What precision do you need? You could probably do a pre-alignment, selecting a suitable area around each curve and do the cross-correlation on much smaller 2D arrays. It might also be sufficient to map the data with less resolution when scaling the x and y into the array indices. I am pretty sure you could easily gain more than a factor of 10. 🙂 try it!

I have some other ideas, bit it's bed time 😉

[alexderjuengere]

---

@altenbach wrote:

---

@alexderjuengere wrote:

---

Interesting idea to do a cross-correlation with the derived 1d arrays!  

---

Of course you could map the data into 2D arrays and do a 2D cross correlation directly. 😉

---

wait a second, I think I have re-invented the wheel ...

if this is the formula for cross-correlation,

the numerator is the product of centered x and centered y, and the denominator is the product of the distances,

but not with regard to 0;0

bur rather to the mean of the template 336;170

dammit.

at least I learned something new.

@ Kahr

good luck in making this quick enough

I think, this is possible, and LabView is capable of doing this

but you will have some work to do...