Abstract: In recent years, under the gradual implementation of mine intelligence-related initiatives, the mining industry has developed towards intelligence and unmanned development. Driverless technology, as an important part of the current intelligent transportation operation system of surface mines, obtains accurate scene geometry information through scene reconstruction and recognition, which is a prerequisite for the application of unmanned transportation vehicles to the operation and production of surface mines. Three-dimensional point cloud data can accurately realize three-dimensional scene reconstruction, and point cloud semantic segmentation can effectively extract the three-dimensional feature information of the road environment in the driving scene, realizing the unmanned driver's accurate recognition of the driving environment area. Compared with urban structured roads, unstructured roads in open pit mining scenarios are characterized by fuzzy road and terrain boundary features, no obvious road edges, and large spatial 3D coordinate spans. In order to solve the problems of small size of publicly available unstructured road datasets, uneven sample distribution, and low segmentation accuracy of mainstream point cloud semantic segmentation algorithms for unstructured roads, a semantic segmentation method for unstructured roads in open-pit mine scenarios is proposed by constructing an open-pit mine point cloud dataset through 3D point cloud reconstruction, as well as by optimizing and improving the PointNet++ algorithm. Based on SFM and MVS algorithms for dense point cloud reconstruction of captured multi-viewpoint images, and optimization and improvement of PointNet++, MLP, channel attention mechanism, and self-attention mechanism based on point attention are introduced, and a point cloud semantic segmentation model for unstructured roads in open-pit mines is designed. In order to verify the effectiveness of the method, 2641 sets of samples of open pit mine point cloud data were constructed based on the S3DIS dataset format for conversion, segmentation and data enhancement. The experimental results of model training show that the improved algorithm for segmentation test has a 4.9% increase in mIoU compared to PointNet++, and the segmentation performance is good. Compared with other point cloud segmentation networks, this network model is more suitable for unstructured roads in open pit mine scenarios, and can meet the requirements of mine transportation unmanned vehicles for driving in feasible domains, providing accurate 3D environment information for subsequent decision-making and planning of unmanned vehicles.