A Robust Big Data Handling Solution for RGB Image Data Set by Indoor UAV-Based Phenotyping System

Images captured by indoor Unmanned Aerial Vehicles (UAV) lack a stable Global Navigation Satellite System (GNSS) or other precise location information unless additional equipment is utilized. Moreover, UAV-captured images tend to have a variety of shooting angles, and the number of shots becomes large, which causes a challenge in terms of computational cost when processing data sets. Therefore, it is necessary to make a method to handle images. The structure from Motion (SfM) method solves this issue. This chapter aims to present the pipeline to make a database about the position of targets within each original image without special equipment. The system operates as follows: Initially, camera positions are estimated using SfM. Subsequently, the areas of each plant or target region are defined as boxes. The vertices of boxes are projected onto each original image. From the result, it could be determined whether the target is within the image. Following projection, the region representing the boxes is filled with white to make mask images. These mask images can significantly reduce computational costs during the construction of dense clouds. This approach also reduces the process associated with training data creation.