In　this　paper,　we　propose　a　framework　called　GridPointNet　that　uses　only　point　clouds　for　3D　object　detection.　There　are　two　important　branches　in　the　research　of　3D　object　detection,　one　is　grid-based　method　and　the　other　is　point-based　method.　These　two　methods　have　their　own　advantages　and　disadvantages,　and　many　previous　works　have　also　tried　to　merge　these　two　methods.　In　this　paper,　we　also　use　this　idea　to　integrate　the　two　methods　into　a　object　detection　framework.　Specifically,　we　use　a　grid-based　method　to　divide　the　point　cloud　into　voxels,　and　then　use　3D　sparse　convolution　to　learn　voxel　features　from　the　point　cloud　and　generate　3D　region　proposals.　At　the　same　time,　we　use　an　improved　point　cloud　sampling　method　to　sample　a　small　number　of　key　points　from　the　entire　scene,　and　then　aggregate　multi-level　voxel　features　based　on　the　point　method.　After　getting　all　the　features,　a　method　called　RoI-Grid　pooling　is　used　to　aggregate　the　features.　Finally,　the　size　and　position　of　the　frame　are　obtained　through　3D　proposal　refinement.　Our　experiments　on　the　KITTI　dataset　show　that　our　framework　has　superior　performance.　©　2022,　Springer　Nature　Singapore　Pte　Ltd.