To improve the efficiency of coal mining and ensure production safety, the construction of intelligent coal mines is being vigorously promoted. With the rapid development of three-dimensional (3D) ranging technology, lidar can acquire high-density 3D laser point clouds in a short time. Semantic segmentation of 3D laser point cloud data can provide accurate environment perception for unmanned mining trucks and realize autonomous operation of those trucks. Point cloud data are characterized by disorder, unstructuredness, and sparsity, which brings difficulties to point cloud data processing and data analysis. To fully learn the local geometric features and the contextual information of mining point clouds and improve the accuracy of semantic segmentation, we propose a new local feature enhancement-based semantic segmentation network called LFE-Net for large-scale laser point clouds in mines.

The LFE-Net input is N×6 mining point cloud. Each point goes through a shared multilayer perceptron (MLP) to obtain eight dimensions as inputs to the encoder-decoder. The encoder-decoder consists of five encoding layers and five decoding layers. Each encoding layer consists of a dilated residual block and a random downsampling operation. The dilated residual block consists of two local feature extraction modules and two local feature aggregation modules, and it is used to enlarge the receptive field of each point. The local feature extraction module learns the spatial position information and surface orientation change information of a given point and its neighborhood points, and utilizes the spatial distance weights to enhance the semantic features of neighborhood points. The local feature aggregation module utilizes mixed pooling to aggregate the features of neighborhood points. Subsequently, we downsample the point cloud and reserve 1/4 of the point cloud for each downsampling. Each decoding layer contains an upsampling operation and an MLP. The decoding layers use upsampling operation to continuously restore the spatial resolution of point cloud features and connect with the intermediate features generated by the encoding layers through skip connections. Finally, the semantic segmentation results are output through three fully connected layers and a dropout layer. Moreover, to alleviate the problem of sample imbalance, a weighted cross-entropy loss function is adopted to make the network pay more attention to small sample classes to improve the accuracy of semantic segmentation.

To conduct experiments on semantic segmentation of the mining point clouds, we utilize a mining truck equipped with lidar sensors to collect laser point cloud data from the open pit mine. We add real semantic labels to the collected point cloud data and create a 3D laser point cloud semantic segmentation dataset of the mine. To fully evaluate the effectiveness of the LFE-Net, we compare its experimental results with other large-scale point cloud semantic segmentation methods by using the mining dataset. The overall accuracy (OA) and the mean intersection over union (mIoU) of the LFE-Net are 97.92% and 87.578%, respectively, which are higher than those of other methods. Additionally, we conduct ablation experiments on the local feature extraction module and the local feature aggregation module to verify the effectiveness of each module.

To fully learn the local geometric features and the contextual information of mining point clouds and improve the accuracy of semantic segmentation, we propose a new local feature enhancement-based semantic segmentation network LFE-Net for large-scale laser point clouds in mines. The main technical contributions of this paper are given as follows. 1) To solve the problem of difficulty in extracting the features of mining trucks, we propose a spatial position encoding unit in this paper. This unit encodes spatial position information of the point clouds and utilizes spatial distance weights to enhance the semantic features of neighborhood points to improve the segmentation accuracy of the algorithm. 2) To solve the problem of difficulty in extracting ground edge features, we propose a spatial normal vector encoding unit. By adding normal vector information, the network can enhance the perception ability of geometric structure features and improve the accuracy of ground edge segmentation. 3) To solve the problem of feature loss caused by max pooling, we propose a mixed pooling. The mixed pooling consists of max pooling and attention pooling, which enriches local features. The LFE-Net is tested on the mining dataset and excellent performance is achieved, with an OA of 97.92% and an mIoU of 87.578%. These experimental results validate the effectiveness of LFE-Net. The network proposed in this paper provides a theoretical basis for the application of unmanned mining trucks in open pit mines, which has significant implications for the unmanned operation of mining trucks.