Semantically grounded explainability for pedestrian detection using anatomical proportions and grid-based visualization