Hmm. You may be mixing up your game (control) logic with your rendering (view) logic.
There are a few ways:
- Grid-placement checks using the grid representing your world in your game logic. In the link you posted, this would be the abstract 2D (or 3D) grid you store in memory as array, that represents the city blocks. This would help in eg. a Sim City type game where you only need to know, "Can I put this here?" Useful for placement, but not so much for analog movement -- good for Chess, but not for PacMan.
- Logical grid checks as axis-aligned boxes PLUS circular collision checks. Here you can use your logical grid as a set of boxes which you can very easily check against one another; but you can also check them against circles representing mobile entities in your game eg. the player, cars on the streets, or whatever. This is simple to implement and very fast because no trig is required really, the only thing you use is overlap checks on the x and y axis (bounding box checks) and pythagoras for radius checks, and point-in-square checks.
- You may want pixel-perfect collision detection. If you do, do a search on it, you will find stuff on here, on gamedev.net forums, on gpwiki.org, all over in fact. It involves checking which pixels of one sprite overlap which pixels of another sprite, in a given rectangular region that represents the overlapping area of the 2 square bitmaps which constitute those sprites. Consider this only if you need a high degree of accuracy, for example in an action game with large, relatively slow moving sprites this might be useful. If your sprites are small and fast moving, check out option 2 -- it will be good enough!
- Use a physics engine like Box2D for collisions if you need really fast collision detection for a great many objects and once, i.e. most of your objects are not static.
If you use approach 1 or 2 or 4, then how you end up rendering after doing the logic I've outlined, is your choice entirely. It could be isometric, it could be top down, side on, full 3D, whatever. But generally there is a distinct difference between your actual world (which is just data) and how it is represented (which is usually just pixels). So don't let the fact that your world is viewed isometrically, confuse you -- it's still just a plain old grid. Do the collision checks on that grid, then worry about whether its isometric or not.