One　of　the　characteristics　of　outdoor　scene　point　clouds　is　their　large　quantity,　so　it　demands　substantial　computational　resources　for　processing.　Sampling　thus　plays　a　critical　role　in　efficient　processing.　Most　existing　methods　overlook　scene　and　task-specific　characteristics,　relying　solely　on　global　point　distribution.　To　address　this,　we　propose　an　adaptive　downsampling　strategy　for　large-scale　outdoor　light　detection　and　ranging　(LiDAR)　point　cloud　registration.　By　statistically　analyzing　semantic　labels,　we　separate　foreground　and　background　point　clouds,　recognizing　that　background　categories　may　vary　across　scenes.　We　then　sample　high-curvature　points　from　the　background　and　contour　points　from　the　foreground　to　preserve　discriminative　spatial　distribution　features.　Extensive　experiments　on　outdoor　datasets　demonstrate　that　our　method　achieves　comparable　performance　to　state-of-the-art methods.　©　2025　The　Author(s).　Electronics　Letters　published　by　John　Wiley　&　Sons　Ltd　on　behalf　of　The　Institution　of　Engineering　and　Technology.