Current Science

P. S. Singh ¹, Mayuri Sharma ², Victor Saikhom ¹, Dibyajyoti Chutia ¹, Chirag Gupta ¹, Avinash Chouhan ¹, P. L. N. Raju ¹

Affiliations
1 North Eastern Space Applications Centre, Department of Space, Government of India, Umiam 793 103, IN
2 Department of Computer Science, Assam Don Bosco University, Guwahati 781 017, IN

Abstract

There has been increasing popularity in large scale mapping for deriving 3D surface and elevation models of earth and building structures. The techniques of computer vision comprising feature detections and matching and photogrammetry play an important role in deriving near accurate 3D reconstruction of scenes from 2D images. Since the images captured by the unmanned aerial vehicle (UAVs) are of high resolution, there is need for more sophisticated processing and analysis of the imagery to generate 3D models and other useful imagery products. The open source softwares are excellent tools for research and can be modified or changed to suit our model, as specific or combinations of algorithms behave differently based on the nature of UAV image scene to be processed. Though many algorithms are available for performing feature extractions from images, few studies have been carried out to identify suitable detector algorithms to be used based on the nature of image or scene that the UAV captures. An attempt has been made to understand and analyse the suitability of feature detection and descriptor algorithms for different scene types. This article also describes the popular technique called structure from motion process pipeline for sequential processing of UAV images with high overlapping, which involves the estimation of 3D point clouds from the keypoint correspondences. The relative accuracy of the 3D point cloud derived from our approach is comparable with similar output from other state-of-the-art UAV processing systems and is found to match with high precision.

Keywords

3D Reconstruction, Open Source, Point Clouds, Remote Sensing, Structure from Motion, Unmanned Aerial Vehicle.

Full Text

   

Puyam S. Singh ¹, Iainehborlang M. Nongsiej ², Valarie Marboh ², Dibyajyoti Chutia ¹, Victor Saikhom ¹, S. P. Aggarwal ¹

Affiliations
1 North Eastern Space Applications Centre, Department of Space, Government of India, Umiam 793 103, IN
2 Department of Computer Science, St Anthony’s College, Shillong 793 001, IN

Abstract

There are challenges in performing 3D scene understanding on point clouds derived from drone images as these data are highly unstructured with no neighbouring information, highly redundant making the processing difficult and time-consuming and have variable density making it difficult to group and segment them. For proper scene understanding, these point clouds need to be segmented and classified into different groups representing similar characteristics. The approaches for segmentation differ based on the distinctiveness of each data product. Although newer machine learning-based approaches work well, they need large amounts of standardized labelled data which in turn require extensive resources and human intervention to obtain good results. Considering these, we have proposed a hybrid clustering-based hierarchical model for effective segmentation of dense 3D point cloud. We have applied the model to local data having a mix of man-made and natural vegetation with variable topography. The combination of RANSAC, DBSCAN and Euclidean method of cluster extraction proved to be useful for precise segmentation and classification of point clouds. The performance of the model has been assessed using Davies–Bouldin dbIndex-based intrinsic measures. The hybrid approach is able to segment 91% of the point clouds precisely compared to the conventional one-step clustering approach.

Keywords

Clustering, Drone Images, Hierarchical Model, Three-Dimensional Point Cloud, Segmentation.

Full Text