Human Pose Estimation Based on Multi-Level Feature Fusion

We propose a human pose estimation algorithm based on multi-level feature fusion module. To enhance information transmission and feature fusion among different feature levels, we design a multi-level feature fusion structure. Further more, we use a multi-level feature fusion module based on a parallel multi-level feature fusion structure to improve the estimation network's ability to process multi-scale targets. The new network, built upon stacked multi-level feature fusion modules, exhibits strong capabilities in learning semantic information, extracting and fusing multi-scale features. When applied to human pose estimation tasks, it can effectively generate high-precision, high-resolution heatmaps for predicting human key points. This algorithm improves the AP(Average Precision) by 1.3 compared to Simple Baselines on the COCO dataset and enhances the [email protected] metric by 0.2 compared to Simple Baselines on the MPII dataset. The experiments demonstrate the effectiveness of the proposed algorithm.