Enhancing PET for crowd counting with adaptive thresholding and polar kernel convolution

Crowd counting and localization in dense scenes face dual challenges: drastic density variations and insufficient local feature modeling. Existing methods typically rely on fixed confidence thresholds and regular convolutional operations, which often lead to high false detection rates in crowded areas and inadequate feature discriminability. To address these issues, this paper proposes an improved framework based on dynamic feature perception. First, a adaptive threshold generation mechanism is introduced to achieve density-adaptive decision-making during decoding, significantly enhancing the model's sensitivity to varying crowd densities. Second, a polar coordinate sampling strategy is incorporated in local regions to strengthen fine-grained feature modeling for dense areas. Experimental results on public datasets demonstrate that our method achieves significant improvements in counting accuracy, outperforming state-of-the-art approaches. The proposed modules are lightweight and easy to integrate, offering a novel solution for crowd analysis tasks in complex real-world environments.