hm...
in the original version he can collide with the outer walls afaik -
how is that done? just by checking x and z value against some value of the wall? Probably.

hm...
I suggest you load an black and white image which you draw in another program and save there.
It is exactly like your maze as a ground plan (seen from above).
Now before taking a step you can check the current x and z on the image plus the next simulated step (as x and y of course and via get() get the color of that pixel:

• When it's black, it's a wall, so collision.
• If it's white, you can go on.

So the trick is to have your maze painted as 3D but to have the same data in an invisible image and check for collisions in the image via get():
http://processing.org/reference/get_.html

you see the idea here:
https://forum.processing.org/topic/my-first-game#25080000001401009


Another principle would be
to save all walls as lines and check them.
This is done here, just let away the drawing of the lines:

http://forum.processing.org/topic/maze-like-game-easing-and-collision-troubles#25080000001420151

Long story short: I tried using both of your examples, but I failed.  Eventually I started trying to get() the pixel color according to my position in the 3D space (versus an imaginary image).  The following lines of code worked for me: