What to use instead of Lukas-Kanade for robust, real time tracking?

I'm currently using LKT tracking for a project of mine, it's 120 fps, so each frame is quite similar to the one before, biggest issues are the slow speed and inability to handle occlusion. I dislike ML based approaches, because they're quite slow. Is there something that can be run on a 600x600 image in <5ms on GPU CUDA?

computervision fps mine project robust something speed tracking