Effects of viewing angle and field of view on detection, tracking, and counting of pine seedlings towards automated forest nursery inventory
Year: 2025
Type: article
Abstract: The current inventory practice in bareroot forest nurseries relies on manually counting tree seedlings in randomly sampled plots to estimate the stock for each seed lot. This method is labor-intensive, time-consuming, and susceptible to human error. Recent advances in deep learning-based object detection and efficient tracking algorithms have enabled automated object counting in video data across various domains, including crop seedling counting in agriculture. This study investigates the effects of viewing angle (VA) and field of view (FoV) on detection, tracking, and counting early-stage pine seedlings in nadir-view videos using a detect-and-track approach. We evaluated the performance of YOLOv8–10 models in conjunction with three multi-object tracking (MOT) algorithms (SORT, ByteTrack, and BoT-SORT) on a custom MOT dataset comprising an average of 153 seedlings per frame and totaling 166,440 seedlings. Detection results and statistical tests showed that increasing horizontal VA reduces the intersection over union (IoU) of seedling detections, primarily due to the perspective differences introduced by oblique viewing angles. MOT evaluations further demonstrated that BoT-SORT consistently delivered high counting accuracy when the vertical FoV encompassed at least the entire seedling. In contrast, ByteTrack and SORT exhibited significantly lower performance, producing reasonable counting accuracy only when the vertical FoV was sufficiently large. The superior performance of BoT-SORT is attributed to its camera motion compensation, which effectively reduces identity switches and tracking failures in scenes involving stationary yet overlapping seedlings. Notably, BoT-SORT achieved 100 % counting accuracy under a 20° horizontal VA across YOLO model sizes. Furthermore, larger YOLO models showed greater robustness to increases in horizontal VA. These findings provide valuable guidance for optimizing camera configurations and model selection towards the development of a real-time inventory systems for precision forest nursery management.
Source: Smart Agricultural Technology
Authors: Ashish Reddy Mulaka, Rafael Bidese, Yin Bao
Institutions: University of Delaware, Corteva (United States), Auburn University
Language: English
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
SDG: Zero hunger
Open Access status: gold
Estimated APC paid: $1,770
Award: G8559633537
OpenAlex ID: W4409385035