Abstract:

Crowd counting takes video surveillance data as input and can be applied to the construction of city digital twin platforms, virtual city modeling and smart city management, etc. However, when there are data domain differences between the application scenario and training scenario, counting performance often significantly decreases. A cross-domain crowd counting model based on frequency domain enhancement is proposed. To alleviate the distribution differences between domains, a frequency domain feature enhancement module and a domain invariant frequency domain adapter module are constructed: the former uses discrete cosine transform to extract key statistical features to enhance spatial representation ability, while the latter decomposes amplitude and phase spectra based on fast Fourier transform and suppresses domain specific amplitude information through attention mechanism to extract domain invariant features. To cope with complex background interference, a parallel dual attention module is designed to focus on the foreground region. To address the challenges posed by drastic scale changes, a multi-scale feature aggregation module is proposed to achievethe fusion of decoder features at different scales, thereby enhancing the model robustness. Cross-domain simulation experiments are conducted on four public crowd counting datasets, and the results showed that the proposed method achieved the lowest counting error in most challenging scenarios, outperforming current mainstream methods and providing effective supports for robust and high-precision crowd dynamic simulation.

Key words: crowd counting, frequency-domain modeling, domain generalization, simulation input modeling, smart city