Laser line scanning technology mounted on underwater vehicles has been widely used for high-precision 3D imaging of seafloor topography and objects. However, target recognition based on line-scan 3D imaging has rarely been reported. Inertial navigation system data are used to perform pose correction for underwater laser line-scanning point clouds. A geometry-semantic fusion recognition framework is then developed by integrating the YOLOv5 network. This framework enables accurate detection and classification of underwater targets. Field experiments and validation were conducted using a self-developed underwater laser line-scanning 3D imaging system mounted on the manned submersible 'Fendouzhe' at a depth of 5 007 m. The sea-trial results demonstrate that, after pose correction, the system can achieve accurate 3D imaging of seafloor targets; for the point cloud of a built-in calibration target, the corrected error can be controlled within 2 cm; and after pose correction, the recognition rate for preset 3D targets reaches 100%. This work provides a new reference for seafloor 3D mapping and target recognition.