Research on Enhanced HRNet for Human Pose Estimation in Crowded Scenes

Abstract

In complex environments such as pedestrian recognition, abnormal behavior detection, and human-computer interaction, high-precision keypoint extraction is the core challenge for bridging low-level visual perception and high level semantic understanding. This paper takes HRNet as the baseline, whose parallel multi-resolution structure maintains high-resolution representations throughout the network, compensating for the maccuracy loss caused by encoder decoder downsampling. However, the native architecture suffers from two limitations in crowded scenes: static convolutional kernels with fixed receptive fields cannot adapt to scale and orientation variations, and feature fusion lacks attention mechanisms and global topological constraints. To address these issues, we propose a deeply optimized HRNet skeleton extraction algorithm with a Hierarchical Attention Mechanism (HAM ) employing dynamic multi-scale patches and parallel feature fusion, and a Dynamic Depthwise Separable Convolution Module(D-DSCM) combining Inception structures with adaptive weight allocation.