Collision　accidents　involving　heavy　construction　machinery　remain　a　leading　cause　of　injuries　in　the　building　industry,　making　real-time　detection　of　site　equipment　and　workers　imperative.　To　address　the　challenges　of　multi-scale　object　detection　in　complex　construction　environments,　this　study　proposes　MSP-YOLO,　an　improved　object　detection　framework　based　on　YOLOv12n.　The　proposed　model　introduces　a　small　object　enhancement　pyramid　(SOEP),　which　incorporates　high-resolution　features　from　shallow　layers　into　the　feature　pyramid　through　SPDConv　and　CSP_Omnikernel,　thereby　enhancing　fine-grained　representation　and　alleviating　the　degradation　of　small　targets.　A　lightweight　multi-axis　gated　coordinate　attention　(MGCA)　mechanism　is　embedded　within　the　neck　to　refine　spatial　feature　encoding　along　height,　width,　and　channel　dimensions,　facilitating　robust　localization　under　occlusion　and　background　clutter.　Additionally,　a　dynamic　selective　weighted　fusion　(DSWFusion)　module　connects　the　backbone　and　neck　via　hierarchical　skip　connections　and　integrates　a　pixel-augmented　convolution　and　attention　mechanism　(PACAM)　with　dynamic　weighting,　promoting　effective　multi-scale　feature　fusion.　These　architectural　components　jointly　improve　the　detection　of　both　small　and　large　objects　while　maintaining　real-time　performance.　Experimental　results　on　our　self-built　construction　site　dataset　demonstrate　that　MSP-YOLO　achieves　64.2%　AP　and　82.5%　AP50　on　the　test　set,　surpassing　YOLOv12n　by　4.5　and　3.7%,　respectively.　It　also　shows　notable　improvements　in　detecting　small　objects　(APs:　+5.3%)　and　medium-sized　objects　(APm:　+8.0%).　Furthermore,　MSP-YOLO　delivers　real-time　performance　with　304.3　FPS　on　an　NVIDIA　RTX　4090,　validating　the　effectiveness　of　our　model.