TY - GEN
T1 - Determining Dendrometry Using Drone Scouting, Convolutional Neural Networks and Point Clouds
AU - Jensen, Kim
AU - Krogh, Oskar Kondrup
AU - Jørgensen, Marius Willemoes
AU - Lehotský, Daniel
AU - Andersen, Anton Bock
AU - Porqueras, Ernest
AU - Søndergaard, Jens Aksel S.
AU - Gade, Rikke
PY - 2021
Y1 - 2021
N2 - This paper presents a solution for mapping the location of trees in an orchard and estimating the dendrometric data of the trees. The combined solution consists of a mapping and navigation algorithm, which allows for autonomous data collection at an orchard with a regular rectangular layout, and data processing for tree detection and dendrometric data estimation. The data collection is done using an Intel RealSense D435i camera, which can obtain both RGB and depth data. The paper presents a comparison between the performance of point cloud processing (PCP) and convolutional neural networks (CNNs) on RGB data for tree detection and dendrometric data estimation. The YOLOv3 CNN achieved a mAP 50 of 63.53% with 65.5 FPS and a mean error of 20.6 cm in height estimation. Point cloud processing achieved a precision of 76.72% with 2.1 FPS and a mean error of 20.4 cm in height estimation. In conclusion, this work shows that point cloud processing shows comparable results to convolutional neural networks for height estimation, but trades off processing time for better precision in detection.
AB - This paper presents a solution for mapping the location of trees in an orchard and estimating the dendrometric data of the trees. The combined solution consists of a mapping and navigation algorithm, which allows for autonomous data collection at an orchard with a regular rectangular layout, and data processing for tree detection and dendrometric data estimation. The data collection is done using an Intel RealSense D435i camera, which can obtain both RGB and depth data. The paper presents a comparison between the performance of point cloud processing (PCP) and convolutional neural networks (CNNs) on RGB data for tree detection and dendrometric data estimation. The YOLOv3 CNN achieved a mAP 50 of 63.53% with 65.5 FPS and a mean error of 20.6 cm in height estimation. Point cloud processing achieved a precision of 76.72% with 2.1 FPS and a mean error of 20.4 cm in height estimation. In conclusion, this work shows that point cloud processing shows comparable results to convolutional neural networks for height estimation, but trades off processing time for better precision in detection.
U2 - 10.1109/CVPRW53098.2021.00326
DO - 10.1109/CVPRW53098.2021.00326
M3 - Article in proceeding
SN - 978-1-6654-4900-7
T3 - IEEE Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)
SP - 2906
EP - 2914
BT - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops (CVPRW)
PB - IEEE (Institute of Electrical and Electronics Engineers)
T2 - 2021 IEEE/CVF Conference on Computer Vision and Pattern Recognition Workshops, CVPRW 2021
Y2 - 19 June 2021 through 25 June 2021
ER -