In　the　field　of　video　surveillance,　the　use　of　artificial　intelligence　to　estimate　crowd　density　in　public　places　has　been　a　popular　study.　In　order　to　improve　the　accuracy　of　crowd　density　estimates,　a　multi-scale　convolution　neural　network　structure　is　proposed.　And　the　feature　fusion　of　different　receptive　field　information　is　performed　by　using　multi-column　convolution　network,　and　the　hierarchical　semantic　information　with　different　feature　maps　at　different　resolutions　is　merged　to　generate　a　crowd　density　map　with　higher　quality.　The　experiment　was　tested　on　the　Shanghaitech　dataset,　UCF_CC_50　dataset,　and　WorldExpo＇10　dataset　with　mean　absolute　error　(MAE)　and　mean　square　error　(MSE)　as　the　evaluation　criteria.　The　results　show　that　the　new　network　model　reduce　the　value　MAE　and　MSE,　improving　the　accuracy　of　crowd　density　estimation.