difference?
* Has anyone here managed to push a CPU-only tracker into the **100–300+ FPS** range on Raspberry Pi-class hardware? If so, what lessons did you learn that aren't obvious from reading papers?

# Future directions I'm considering

Beyond getting the core tracker running efficiently, some areas I'd like to explore include:

* DSST-based scale estimation.
* Lightweight re-detection and automatic target re-acquisition.
* More robust confidence estimation beyond PSR.
* Hybrid detector–tracker pipelines that combine fast tracking with occasional detection.
* FFT optimization, cache-aware memory layouts, and ARM/NEON-specific performance tuning.
* General techniques for squeezing the maximum performance out of embedded CPU-only vision systems.

I'm not looking for someone to redesign the project or suggest replacing it with deep learning. My goal is to understand where the real bottlenecks are and learn from people who've already built or optimized similar systems before I spend weeks optimizing the wrong component.

If you've worked on anything similar, or achieved high frame rates with classical tracking methods, I’d love to hear about your experience, benchmark results, profiling insights, or even things that *didn't* work. Thanks in advance!

https://redd.it/1ueepo7
@r_cpp