The point cloud data of the container ship cargo hold obtained by the 3D scanner is characterized by high accuracy and large data volume. To address the challenges of processing large-scale point cloud data with existing deep learning networks, this study employs the encoding structure of RandLA-Net as the backbone network to achieve effective component recognition in large-scale cargo hold point clouds. In the segmentation task, the original RandLA-Net is utilized as the backbone network. To address the limitations of single-feature aggregation in the feature aggregation module, a bilaterally offset-based feature aggregation module is introduced for the recognition task. A point cloud dataset of the container ship cargo hold was constructed and used to train the proposed network. Experimental results show that the model achieved an mAcc of 78.37% on the cargo hold point cloud dataset, effectively segmenting the rail track point cloud, demonstrating that the proposed method exhibits strong segmentation performance in large-scale point cloud segmentation tasks.