Kazuo Oda

Automatic registration of MLS point clouds and SfM meshes of urban area

Abstract Recent advances in 3D scanning technologies allow us to acquire accurate and dense 3D scan data of large-scale environments efficiently. Currently, there are various methods for acquiring large-scale 3D scan data, such as Mobile Laser Scanning (MLS), Airborne Laser Scanning, Terrestrial Laser Scanning, photogrammetry and Structure from Motion (SfM). Especially, MLS is useful to acquire dense point clouds of road and road-side objects, and SfM is a powerful technique to reconstruct meshes with textures from a set of digital images. In this research, a registration method of point clouds from vehicle-based MLS (MLS point cloud), and textured meshes from the SfM of aerial photographs (SfM mesh), is proposed for creating high-quality surface models of urban areas by combining them. In general, SfM mesh has non-scale information; therefore, scale, position, and orientation of the SfM mesh are adjusted in the registration process. In our method, first, 2D feature points are extracted from both SfM mesh and MLS point cloud. This process consists of ground- and building-plane extraction by region growing, random sample consensus and least square method, vertical edge extraction by detecting intersections between the planes, and feature point extraction by intersection tests between the ground plane and the edges. Then, the corresponding feature points between the MLS point cloud and the SfM mesh are searched efficiently, using similarity invariant features and hashing. Next, the coordinate transformation is applied to the SfM mesh so that the ground planes and corresponding feature points are adjusted. Finally, scaling Iterative Closest Point algorithm is applied for accurate registration. Experimental results for three data-sets show that our method is effective for the registration of SfM mesh and MLS point cloud of urban areas including buildings.

Segmentation of Lidar Intensity using Weighted Fusion based on Appropriate Region Size.

We propose a semantic segmentation method for LiDAR intensity images obtained by Mobile Mapping System (MMS). Conventional segmentation method could give high pixel-wise accuracy but the accuracy of small objects is quite low. We solve this issue by using the weighted fusion of multi-scale inputs because each class has the most effective scale that small object class gives higher accuracy for small input size than large input size. In experiments, we use 36 LIDAR intensity images with ground truth labels. We divide 36 images into 28 training images and 8 test images. Our proposed method gain 87.41% on class average accuracy, and it is 5% higher than conventional method. We demonstrated that the weighted fusion of multi-scale inputs is effective to improve the segmentation accuracy of small objects.

Semantic Segmentation in Red Relief Image Map by UX-Net.

This paper proposes a semantic segmentation method in Red Relief Image Map which a kind of aerial laser image. We modify the U-Net by adding the paths between convolutional layer and deconvolutional layer with different resolution. By using the feature maps obtained at different layers, the segmentation accuracy is improved. We compare the segmentation accuracy of the proposed UX-Net with the original U-net. Our proposed method improved class-average accuracy in comparison with the U-Net.

Road Detection from Satellite Images by Improving U-Net with Difference of Features.

In this paper, we propose a road detection method from satellite images by improving the U-Net using the difference of feature maps. U-Net has connections between convolutional layers and deconvolutional layers and concatenates feature maps at convolutional layer with those at deconvolutional layer. Here we introduce the difference of feature maps instead of the concatenation of feature maps. We evaluate our proposed method on road detection problem. Our proposed method obtained significant improvements in comparison

Segmentation of LiDAR Intensity Using CNN Feature Based on Weighted Voting

We propose an image labeling method for LiDAR intensity image obtained by Mobile Mapping System (MMS). Conventional segmentation method using CNN and KNN could give high accuracy but the accuracies of objects with small area are much lower than other classes with large area. We solve this issue by using voting cost. The first cost is determined from a local region. Another cost is determined from surrounding regions of the local region. Those costs become large when labeling result corresponds to class label of the region. In experiments, we use 36 LIDAR intensity images with ground truth labels. We divide 36 images into training (28 images) and test sets (8 images). We use class average accuracy as evaluation measures. Our proposed method gain 84.75% on class average accuracy, and it is 9.22% higher than our conventional method. We demonstrated that the proposed costs are effective to improve the accuracy.