An innovative segmentation method with multi-feature fusion for 3D point cloud

<p><strong>Abstract.</strong> Structural disaster damage detection and characterisation is one of the oldest remote sensing challenges, and the utility of virtually every type of active and passive sensor deployed on various air- and spaceborne platforms has been assessed. The proliferation and growing sophistication of UAV in recent years has opened up many new opportunities for damage mapping, due to the high spatial resolution, the resulting stereo images and derivatives, and the flexibility of the platform. We have addressed the problem in the context of two European research projects, RECONASS and INACHUS. In this paper we synthesize and evaluate the progress of 6 years of research focused on advanced image analysis that was driven by progress in computer vision, photogrammetry and machine learning, but also by constraints imposed by the needs of first responder and other civil protection end users. The projects focused on damage to individual buildings caused by seismic activity but also explosions, and our work centred on the processing of 3D point cloud information acquired from stereo imagery. Initially focusing on the development of both supervised and unsupervised damage detection methods built on advanced texture features and basic classifiers such as Support Vector Machine and Random Forest, the work moved on to the use of deep learning. In particular the coupling of image-derived features and 3D point cloud information in a Convolutional Neural Network (CNN) proved successful in detecting also subtle damage features. In addition to the detection of standard rubble and debris, CNN-based methods were developed to detect typical façade damage indicators, such as cracks and spalling, including with a focus on multi-temporal and multi-scale feature fusion. We further developed a processing pipeline and mobile app to facilitate near-real time damage mapping. The solutions were tested in a number of pilot experiments and evaluated by a variety of stakeholders.</p>

Download Full-text

GRAPH NEURAL NETWORK BASED MULTI-FEATURE FUSION FOR BUILDING CHANGE DETECTION

ISPRS - International Archives of the Photogrammetry, Remote Sensing and Spatial Information Sciences ◽

10.5194/isprs-archives-xliii-b3-2021-377-2021 ◽

2021 ◽

Vol XLIII-B3-2021 ◽

pp. 377-382

Author(s):

W. Yuan ◽

X. Yuan ◽

Z. Fan ◽

Z. Guo ◽

X. Shi ◽

...

Keyword(s):

Neural Network ◽

Remote Sensing ◽

Change Detection ◽

Point Cloud ◽

Feature Fusion ◽

Experimental Comparison ◽

3D Point Cloud ◽

Remote Sensing Images ◽

Illumination Variations ◽

Multi Temporal

Abstract. Building Change Detection (BCD) via multi-temporal remote sensing images is essential for various applications such as urban monitoring, urban planning, and disaster assessment. However, most building change detection approaches only extract features from different kinds of remote sensing images for change index determination, which can not determine the insignificant changes of small buildings. Given co-registered multi-temporal remote sensing images, the illumination variations and misregistration errors always lead to inaccurate change detection results. This study investigates the applicability of multi-feature fusion from both directly extract 2D features from remote sensing images and 3D features extracted by the dense image matching (DIM) generated 3D point cloud for accurate building change index generation. This paper introduces a graph neural network (GNN) based end-to-end learning framework for building change detection. The proposed framework includes feature extraction, feature fusion, and change index prediction. It starts with a pre-trained VGG-16 network as a backend and uses U-net architecture with five layers for feature map extraction. The extracted 2D features and 3D features are utilized as input into GNN based feature fusion parts. In the GNN parts, we introduce a flexible context aggregation mechanism based on attention to address the illumination variations and misregistration errors, enabling the framework to reason about the image-based texture information and depth information introduced by DIM generated 3D point cloud jointly. After that, the GNN generated affinity matrix is utilized for change index determination through a Hungarian algorithm. The experiment conducted on a dataset that covered Setagaya-Ku, Tokyo area, shows that the proposed method generated change map achieved the precision of 0.762 and the F1-score of 0.68 at pixel-level. Compared to traditional image-based change detection methods, our approach learns prior over geometrical structure information from the real 3D world, which robust to the misregistration errors. Compared to CNN based methods, the proposed method learns to fuse 2D and 3D features together to represent more comprehensive information for building change index determination. The experimental comparison results demonstrated that the proposed approach outperforms the traditional methods and CNN based methods.

Download Full-text