Classification and Segmentation of Mining Area Objects in Large-Scale Spares Lidar Point Cloud Using a Novel Rotated Density Network

Abstract

The classification and segmentation of large-scale, sparse, LiDAR point cloud with deep learning are widely used in engineering survey and geoscience. The loose structure and the non-uniform point density are the two major constraints to utilize the sparse point cloud. This paper proposes a lightweight auxiliary network, called the rotated density-based network (RD-Net), and a novel point cloud preprocessing method, Grid Trajectory Box (GT-Box), to solve these problems. The combination of RD-Net and PointNet was used to achieve high-precision 3D classification and segmentation of the sparse point cloud. It emphasizes the importance of the density feature of LiDAR points for 3D object recognition of sparse point cloud. Furthermore, RD-Net plus PointCNN, PointNet, PointCNN, and RD-Net were introduced as comparisons. Public datasets were used to evaluate the performance of the proposed method. The results showed that the RD-Net could significantly improve the performance of sparse point cloud recognition for the coordinate-based network and could improve the classification accuracy to 94% and the segmentation per-accuracy to 70%. Additionally, the results concluded that point-density information has an independent spatial–local correlation and plays an essential role in the process of sparse point cloud recognition.

MLFNet- Point Cloud Semantic Segmentation Convolution Network Based on Multi-Scale Feature Fusion

“…The goal of segmentation [18] is achieved by constructing a point cloud using a relevant point cloud processing platform, which results in many misidentifications. The task of segmentation is affected by the density of the point cloud, so high-precision density segmentation has been adopted [19,20], but the characteristics of sensitive areas were lost. To address this problem, a constrained network of points [21] has been adopted.…”

Methods in the spatial deep learning: current status and future direction

“…Yang et al [69] jointly trained two parallel U-Net based network of 9 layers with lower layers shared between the source domain and target domain and fine-tuned the upper layers of the target domain using pseudo labels from unsupervised learning. Networks built using supervised classifiers such as extreme learning machines and SVM on top of DBN, DBN was first trained unsupervised manner and subsequently fine-tuned with labeled data [70,71]. For change detection, deep learning was used in an end-to-end manner or with features extracted from the input images as input for the deep learning model [70].…”

“…Some papers demonstrated the possibilities of successful transfer learning using synthetic data [70] or using secondary manually labeled data, and the effectiveness of dimensionality reduction of multispectral satellite images for reducing computation load without compromising accuracy [75]. Yan et al [71] focused on data preprocessing and the development of an auxiliary network called rotation density network to simplify as well as improve the segmentation of 3D point clouds of the mining area.…”

“…The segmentation of archeological and cultural heritage sites was performed with a dynamic graph convolution neural network (DGCNN) constructed with several blocks of edge convolutional layers [74]. Yan et al [71] developed an auxiliary network called Rotation Density Network that efficiently extracts structural features based on the density of point cloud and used it in conjunction with point-based networks PointNet or PointCNN to enhance the point cloud segmentation of the mining area.…”

“…Vaihengen and Potsdam datasets were divided into training/test sets with 16/17 and 24/14 tiles. Similarly, 80% used for training and 20% for testing were used in [85] and 75% data as a training sample and 25% as testing in [71]. Braga et al [87], divided the 19,656 samples into 15,122 as a training sample and 4535 images as test samples.…”