In　this　paper,　we　focus　on　the　problem　of　small　and　longrange　object　misses　in　3D　object　detection　on　point　clouds.　We　observed　that　in　challenging　situations,　especially　for　hard　objects　such　as　small　objects,　the　performance　of　the　detector　remains　unsatisfactory.　To　address　these　issues,　this　paper　proposes　a　voxel-based　two-stage　3D　object　detector,　named　DA-TSD,　which　mainly　includes　a　Double　Attention　(DA)　module　and　a　Pyramid　Sampling　(PS)　module.　The　DA　module　comprehensively　considers　point-wise　and　channel-wise　excitation　attention,　which　can　effectively　enhance　the　crucial　information　of　the　object　and　suppress　irrelevant　noise.　In　addition,　the　stacked　DA　module　utilizes　not　only　the　current　level　feature　but　also　the　multi-level　feature　attention.　The　PS　module　provides　cross-layer　feature　mappings　to　obtain　more　comprehensive　feature　representations.　The　experimental　results　on　the　val　set　of　the　KITTI　dataset　demonstrate　the　superiority　and　effectiveness　of　DA-TSD.　DA-TSD　provides　higher　detection　accuracy　while　maintaining　real-time　frame　processing　rate,　running　at　a　speed　of　28.5　FPS　on　an　NVIDIA　GeForce　RTX　3090　Ti　GPU.