Global Context-Aware Feature Extraction and Visible Feature Enhancement for Occlusion-Invariant Pedestrian Detection in Crowded Scenes