That is odd. I know blitz sphere->poly collisions can get really slow when a large(ish) amount of polies are involved. That doesn't really explain why there's such a large discrepency when going around a bend, though. I suppose collisions are actually occuring a lot more when going around a bend - maybe this has an impact on speed, for some reason?
Are your collision pivots situated at the 4 corner wheels of each carriage, and you're just checking for lateral collisions with the side of the track? If so, I'd suggest having a separate, minimal proxy mesh of the track, and check for collisions with that. As you're only checking for lateral collisions(?), your proxy mesh onviously wouldn't need any polies on the top/bottom. Also, you could probably represent the track sides in the proxy mesh by just using a 'chain' of connected tris, instead of quads.
I actually had to implement this very method myself recently, to overcome a very similar problem, although I was only doing one collision check, each frame. It worked pretty well for me.
P.S. I know collisions were 'tweaked' some time after version 1.87, so it may be worth trying the latest version to see if it makes any difference.
Are your collision pivots situated at the 4 corner wheels of each carriage, and you're just checking for lateral collisions with the side of the track? If so, I'd suggest having a separate, minimal proxy mesh of the track, and check for collisions with that. As you're only checking for lateral collisions(?), your proxy mesh onviously wouldn't need any polies on the top/bottom. Also, you could probably represent the track sides in the proxy mesh by just using a 'chain' of connected tris, instead of quads.
I actually had to implement this very method myself recently, to overcome a very similar problem, although I was only doing one collision check, each frame. It worked pretty well for me.
P.S. I know collisions were 'tweaked' some time after version 1.87, so it may be worth trying the latest version to see if it makes any difference.