This paper introduces BFVTModel, an efficient 3D semantic segmentation Model. Multi-modal segmentators, which employ LiDAR and Camera sensors as input, have gained popularity owing to their capacity to leverage the semantic information in image data and complementary details in point cloud data. However, the segmentation model using multi-modal input faces some problems, such as the modal consistency problem and the model complexity problem. In this paper, we propose utilizing the idea of an Occupancy-based multi-view feature combination method to obtain the 3D information of the features of the two modalities independently. We also design a projection structure to obtain the three-axis features in different modalities according to the information, including the main view, side view, and top view features. We construct a module called Feature View Transform (FVT) that combines three-axis plane features with bias. The CNN neural network is used to replace attention mechanisms, reducing the dimensionality of the original three-dimensional space matrix, thereby reducing the parameter count and increasing the model’s speed. We used a bilinear mapping structure to fuse LiDAR and Camera features and complete the 3D semantic segmentation task. We validate the model on the NuScenes dataset and obtain competitive accuracy results while leading the way in efficiency evaluation metrics.