CropFinder: AI-based Detection and Tracking of Crops for Precision Agriculture

Abstract

Individual plant treatment, such as irrigation, fertilization, and pesticide application is pivotal in precision agriculture. Traditional methods, which detect all plants using robotic cameras and apply uniform farming practices, are not cost-effective or eco-friendly. This study introduces efficient and precise detection and tracking methods for individual lettuce plants. Our method automates the annotation of lettuce plants using the Grounded Segment Anything Model (SAM) to train the YOLOv8m detection model, achieving a mean average precision of $87 \%$. For tracking, the study employed ByteTrack and deep observation-centric simple online and real-time tracking algorithms. The latter yielded the best results with an $87 \%$ higher order tracking accuracy due to its deep feature association methods, enabling re-identification capabilities. This algorithm also supports real-time tracking, effectively differentiating and identifying plants even during extended occlusions, such as camera retraction. Further improvements in tracking accuracy were explored using grayscale imagery and varying contrast levels in video streams, with a contrast level of 1.5 achieving an accuracy of $87.8 \%$. Compared to existing methods, this study achieves better detection and tracking performance in agricultural settings.