2018 年 74 巻 4 号 p. 154-161
Recently, structure from motion (SfM), which converts multiple images to a detailed three-dimensional (3D) model, has been used to extract 3D structural information about vegetation. However, multiple still images (e.g., >100 images) are necessary for the 3D reconstruction. This requires multiple shutter releases, but taking many images is time consuming and labor intensive. One possible solution is to take video recordings from which many images can be obtained by dividing the video clips into video frames. However, frames from videos are sometimes blurred owing to camera vibration, which leads to inaccurate construction of the 3D model. Furthermore, their resolution is lower than that of still images, which may lead to inaccurate 3D reconstruction and estimation error of tree trunk diameter, tree height, and the number of trees observable in the 3D images. We propose a method to record videos, remove blurred video frames using machine learning, and construct 3D images. We compare the accuracy of the 3D models reconstructed from video frames with that from the still images. The blurred video frames were classified by a convolutional neural network (CNN) with an accuracy of 97%. The classification to remove these video frames improved the accuracy of the 3D models based on video frames taken at a walking speed of more than 0.5 m/s, which included many blurred ones. There was no significant difference in the accuracy of tree trunk diameter and tree height estimation between the 3D models obtained from the video frames and the still images when using the CNN classification. At a close enough distance (e.g., 20 m), the 3D model reconstructed from video frames was as accurate as the models constructed from still images. Video recording enables effective data collection for SfM, and the present method can be applicable to the 3D reconstruction of trees in various fields.
