Hey, that's really impressive !
Not sure if you're ready to share your code yet or not - but we might be able to get out a little more/better performance (or maybe even just more vehicles for the same runtime footprint) - for example, there are some low-level Nasal optimizations that we can do, because the Nasal code generator is pretty dumb actually. But I haven't done any actual profiling/benchmarking yet, so this would be mainly about suggesting different coding constructs and benchmarking those to see if/how they perform better or worse.
I do think that a simple tile/LOD-scheme based on viewer position, orientation, FOV and groundspeed might be worthwhile to explore (maybe in conjunction with relative speed/movement direction of the generated traffic) - I believe Thorsten's Advandced Weather system was/is using a 3x3 grid of "tiles" to create a cache-based system around the position of the current aircraft - something like that should work also fairly well here I suppose:
http://wiki.flightgear.org/Weather#Weather_tilesTo be honest, even in the current state it's funny to see that your scripted approach works obviously better than the hard-coded random buildings system in the video, because the latter can be seen adding/removing models at runtime, whereas I didn't spot anything weird going on with the generated road traffic.
Regarding wkitty42's suggestion, I believe that this would be mainly a matter of having better 2D vector data, probably including a 3rd (elevation) component -possibly involving OSM data (the background thread approach would also work for pre-computing such tables asynchronously, e.g. by loading such data from disk).
Anyway, I do think it would be a good idea to structure the whole thing as an addon and get this committed to fgaddon (provided it will be GPL?) I am sure you could get Thorsten involved to help with this, because he's been working on the Traffic Shader anyway - and who knows, maybe we can find a way to integrate these two approaches ?