Density-Guided Adaptive Patch Learning for Robust Crowd Counting

Abstract: Accurate crowd counting in real-world scenes re-mains challenging due to severe occlusions, perspective distortion, and large intra-scene density variation. Recent deep learning based approaches typically address these challenges using patch-level learning, where images are divided into fixed grids or randomly cropped patches. These approaches then estimate the count in each patch. However, such fixed partitioning strategies often fail to align with the irregular spatial distribution of crowds. This leads to heterogeneous density patterns within patches where the models fail to produce the accurate count. In this study, we propose a simple, yet effective Density-Guided Adaptive Patch Learning framework for crowd counting. Instead of relying on fixed-size patches, we first obtain a coarse density estimation to capture the global density structure of a scene. Based on this estimate, the image is dynamically partitioned into density-homogeneous regions, where dense areas are represented using smaller patches and sparse regions using larger patches. Each adaptive patch is then processed independently for density estimation, and the resulting predictions are fused to produce the final crowd density map. The proposed framework is model-agnostic and can be seamlessly integrated with existing crowd counting networks without architectural modification. Extensive experiments on benchmark datasets demonstrate that the pro-posed adaptive partitioning consistently achieves lower Mean Absolute Error (MAE) in counting accuracy and localization compared to fixed patch-based baselines, particularly in scenes with strong density variation.

Abdullah N Alhawsawi. “Density-Guided Adaptive Patch Learning for Robust Crowd Counting”. International Journal of Advanced Computer Science and Applications (IJACSA) 17.2 (2026). http://dx.doi.org/10.14569/IJACSA.2026.0170292

@article{Alhawsawi2026,
title = {Density-Guided Adaptive Patch Learning for Robust Crowd Counting},
journal = {International Journal of Advanced Computer Science and Applications},
doi = {10.14569/IJACSA.2026.0170292},
url = {http://dx.doi.org/10.14569/IJACSA.2026.0170292},
year = {2026},
publisher = {The Science and Information Organization},
volume = {17},
number = {2},
author = {Abdullah N Alhawsawi}
}