Enhanced street graph processing of OSM areas

[vesameskanen]

Summary

Street graph generation code, which processes OSM areas, refactorized and optimized.

• Boarding location reference list, which is needed only during graph building, moved away from graph's area group class
• Stop linking with boarding areas is now more accurate, because it uses original platform centroid instead of the centroid of merged adjacent areas. This used to caused strange teleportation hundreds of meters away from platforms.
• Walkable area building optimized: unnecessary self-intersection tests skipped and buggy bookkeeping of existing connections fixed. The slowest phase of street graph building is now over two times faster.
• Visibility edges, which cross multiple areas, are no longer recursively cut into segments. Instead, area properties are accumulated into a single edge.
• Removing duplicate area edges makes street routing faster and more efficient
• Several poorly named classes (and some of their fields) renamed

The images below show stop linking in Kerava railway station. After the changes, a railway stop gets linked to a platform. In the past, a stop link teleported travellers far away from the platforms.

Before:

After:

Issue

Closes #6103

Unit tests

Some existing tests updated to reflect new improvements.

Documentation

Lots of outdated code comments removed and updated.

[leonardehrenfried]

Area doesn't seem to have equals or hashcode. Are the maps using idenity in that case?

[vesameskanen]

I did not even realize this issue, just trusted Java magic. I have to check this.

[vesameskanen]

I guess java uses object address as hash or something like that. I do not see any problem in this.

[leonardehrenfried]

I believe this is intended, but can you explain what is going on?

[vesameskanen]

One of the optimizations was adding a threshold for convexity. 3 points along a straight line were randomly interpreted to define a convex or concave angle, because of inaccuracy of computations. Only true concave points matter in visibility computations. Herrenberg test platform lost one unnecessary fake visibility point which is good.

[leonardehrenfried]

Generally this looks very good and I can confirm that it resolves an issue in one of my data sets.

I understand that you radically simplified the WalkableAreaBuilder. I would love to understand what you did and what you removed. Are you available for a quick meeting about this?

[leonardehrenfried]

Can you pull this commit so that we have a bit of documentation on equals/hashcode leonardehrenfried@f0a94d2

?

[t2gran]

Area doesn't seem to have equals or hashcode. Are the maps using idenity in that case?

OTP 1.x relayed on Object#equals() and Object#hashCode(). This is in some situations much faster than overriding these methods and base the implementation on comparing fields. I recommend to do it explicit so it is clear why you do it:

  /**
   * Object identity is used for eq/hc. In this case this is safe since instances are 
   * created during graph build and is guaranteed to be unique.
   */
  @Override
  public final boolean equals(Object other) {
    return super.equals(other);
  }

  /**
   * See {@link #equals(Object)}
   */
  @Override
  public final int hashCode() {
    return super.hashCode();
  }

Note! Both methods are final.