The position of prestressed ducts has a significant impact on the bearing capacity of prestressed concrete members. A position detection method for ducts based on 3D laser scanning and deep learning is proposed to improve the efficiency and accuracy of ducts position detection. Firstly, a dataset containing prestressed ducts is produced by the 3D laser scanning method. Then, the ducts point cloud is obtained by an efficient large-scale 3D point cloud semantic segmentation network (RandLA-Net). The centroid of the ducts point cloud slice is obtained by using the bracket box midpoint method and circle fitting constraints method. Finally, the centerline of the ducts is fitted by the back-propagation neural network (BP network). The detection results of actual prestressed concrete box girder duct position using this method show that even under the interference of complex surrounding environments such as box girder reinforcing steel skeleton and construction tire frame, this method can obtain the complete prestressed duct position. The maximum detection error is less than ±10 mm, which meets the requirements of relevant construction specifications.