Extend PrivHAR-Bench to Multi-Person Privacy-Preserving HAR

Develop a multi-person extension of the PrivHAR-Bench benchmark that supports privacy-preserving human activity recognition in scenes containing multiple people, rather than selecting only the single largest detected person per frame, and establish standardized evaluation protocols for such multi-person scenarios.

Background

PrivHAR-Bench currently assumes a single-person scenario: the pipeline selects the single largest detected person per frame, and classes involving person-person interactions are excluded. As a result, multi-person scenes (e.g., group activities or crowded environments) are not addressed by the present benchmark.

This constraint limits the applicability of the benchmark to real-world settings where multiple individuals are often present and potentially interact. Extending PrivHAR-Bench to handle multi-person scenes would require representing and transforming multiple regions of interest per frame and defining evaluation protocols that remain privacy-preserving while enabling fair, reproducible comparison across methods.