Depth　information　is　crucial　for　an　autonomous　driving　system　as　it　helps　the　system　understand　the　environment　and　make　decisions.　Most　deep　learning-based　depth　completion　methods　are　primarily　designed　for　high-resolution　lidars　(e.g.　64　scanlines).　However,　when　the　number　of　lidar　scanlines　decreases,　such　as　with　32　scanlines　or　16　scanlines　lidars,　existing　solutions　may　face　challenges　in　reliably　predicting　dense　depth　maps.　To　address　this　issue,　this　paper　proposes　an　effective　framework　based　on　knowledge　distillation,　which　incorporates　mixed-scanline　resolution　training　and　feature-level　fusion　to　train　a　powerful　teacher　network　that　dynamically　fuses　features　from　high-scanline　resolution　and　low-scanline　resolution　inputs.　By　supervising　the　student　network　based　on　the　guidance　of　the　teacher　network,　the　knowledge　from　the　multi-scale　fusion　teacher　network　is　effectively　transferred　to　the　low-scanline　resolution　student　network.　For　the　inference　process,　only　the　student　network　is　utilized.　The　proposed　framework　has　been　applied　to　various　existing　depth　completion　networks.　The　experimental　results　show　the　effectiveness　of　the　proposed　method　by　using　the　KITTI　dataset,　which　shows　that　it　can　serve　as　a　universal　framework　for　depth　completion　tasks.　©　2024　IEEE.