Vectorize line zone

[tc360950]

Description

This is a follow-up to 1205.
In this PR I've vectorized LineZone to achieve better performance.

As previously all three approaches (the one from develop, the one from 1205, and the one from this PR)
have been run on Yolov8m + ByteTrack + LineZone run 10 times on market square video.
This time I've also checked what happens when all detections are duplicated X times (resulting in a larger number of tracked objects).
The total time each LineZone version takes to compute on all frames is (seconds):

Version / Number of duplicates	1 (no duplicates)	2	4	8
develop	11.1	33	77	164
previous PR	0.71	1.7	3.1	5.53
This PR (vectorized version of develop	1.17	2.7	5	8.7

As you can see the one from previous PR always "wins" but I find the difference small enough, especially if you find this version more readable.

Type of change

Please delete options that are not relevant.

• Optimization (non-breaking change which optimizes existing feature)

[SkalskiP]

Hi @tc360950 👋🏻
Considering your results and analyzing the increase in code complexity, I definitely lean towards choosing this PR over #1205.

[LinasKo]

Hi @tc360950,

What's the best way to test this? Is there a Colab we can use?

[LinasKo]

I did some basic verification with this code: https://colab.research.google.com/drive/1rA0wXfaPlGnGKzn6wVA_x01jcX9Dr8Gu?usp=sharing

I see no regressions, after testing with vehicles and people-walking videos.
I ran the updated tests from #1206 and it behaves identically, passing the current tests and failing on V-shape motions.

@tc360950, if you have a better test, let us know 😉

[LinasKo]

Hey @tc360950,

I've submitted some review comments. The changes are very small - if you can finish by Tuesday, there's a very high chance we'd include this in the next release! 🚀

[tc360950]

I did some basic verification with this code: https://colab.research.google.com/drive/1rA0wXfaPlGnGKzn6wVA_x01jcX9Dr8Gu?usp=sharing

I see no regressions, after testing with vehicles and people-walking videos. I ran the updated tests from #1206 and it behaves identically, passing the current tests and failing on V-shape motions.

@tc360950, if you have a better test, let us know 😉

I actually went a little further - I've pasted the vectorized version into a notebook and compared results to the old implementation after every frame:
https://colab.research.google.com/drive/14tPYRcgFo9-WZSppEt7pDB_djRmDQHqG?usp=sharing

[LinasKo]

This is awesome. A very high quality contribution.
I've had a look at the colab - it's indeed a very thorough test.

Thank you very much!