A multimodal point cloud semantic segmentation algorithm assisted by 2D image information was proposed in response to the excessive computational cost introduced by incorporating 2D image information in multimodal semantic segmentation methods, as well as the significant feature discrepancy and fusion difficulty between different modalities. A dual-branch feature extraction network was employed to separately extract point cloud feature and RGB image feature. A feature separation convolution was designed to optimize the extracted 2D image feature in order to reduce the additional computational burden caused by the introduction of RGB images. A point-to-pixel mapping strategy was adopted to transform 3D point cloud feature into a representation compatible with 2D image feature. A multimodal knowledge alignment module was designed. The key feature of point cloud and RGB image was emphasized through a dual-fusion attention mechanism. Then the point cloud feature and RGB image feature were mapped into a unified feature space. The discrepancy between modalities was eliminated, thereby allowing the model to perceive the environment more comprehensively. The proposed algorithm achieved semantic segmentation accuracy of 69.1% and 79.3% on the SemanticKITTI and NuScenes datasets, respectively, outperforming other mainstream algorithms. The experimental results demonstrate that the proposed algorithm effectively improves the accuracy of point cloud semantic segmentation by combining point cloud information with RGB image information and exploiting the complementary advantage of multimodal data.