To meet the challenge that the current semantic segmentation methods of LIDAR point cloud image are difficult to balance the processing speed and accuracy, and takes up huge computing resources, a lightweight algorithm based on codec structure is studied. This study firstly uses spherical projection to project 3D point clouds to two-dimensional planes, then designs a lightweight two-dimensional semantic segmentation network MobileSeg based on MobileNetV3 to segment projection map, and finally projects the segmentation results back to the 3D point cloud. The algorithm reduces the occupation of computing resources by reducing the dimensionality of point cloud images and lightweight backbone networks, and avoids various difficulties faced when using neural networks to directly extract features from sparse point clouds. The average intersectional union ratio (mIOU) of the algorithm on the SemanticKITTI dataset is 51.7%, and the parameters of the MobileSeg-S and MobileSeg-L models are 0.9M and 3.2 M, respectively. The accurate and lightweight point cloud semantic segmentation method has a certain application prospect in the field of automatic driving.