ASHS 2025 Conference

Efficient and accurate counting of plants is critical for nursery inventory management to support yield prediction, sales forecasting, and monitoring. Current practices in nurseries depend heavily on manual methods, which are labor-intensive and prone to errors. Researchers have made efforts in utilizing computer vision and deep learning to address these issues, yet a seamless solution for plant counting and inventory management remains unavailable. Image-based counting systems often struggle with classification accuracy in diverse, real-world scenarios, while tracking plants in videos faces challenges such as identity switches, misclassifications, and varying field conditions, limiting the reliability of existing methods. To overcome these challenges, we developed a cloud-based complete solution specifically for ornamental plant nursery inventory management. Our system introduces a novel tracking algorithm KBTrack, optimized for precision and scalability. At its core is an ensemble deep learning model that combines a transfer learning-based YOLOv11 and a CutMix-enhanced YOLOv11 model for plant detection. The KBTrack was developed on top of ByteTrack multi object tracking framework by adding a layer that compares segmentation masks across multiple frames for long term matching of the objects. This addition addressed identity switches and misclassifications, ensuring accurate plant counting even in complex field conditions. The segmentation capabilities of YOLOv11 are also utilized to generate masks for individual plants, enabling customizable plant-specific quality assessments through an interactive dashboard. The system utilizes GPS to allow users to monitor nursery plant beds on a map making it easier to monitor and track changes and updates across the plant beds. The platform is deployed in cloud with a microservice architecture where users can upload field videos and access results through an intuitive interface designed to ensure scalability. To evaluate the capabilities of the proposed framework, data was collected using an autonomous ground vehicle equipped with an OAK-D Pro camera, capturing 4K resolution videos. Experiments conducted on Azalea and Sunshine plants demonstrated the system's effectiveness, achieving a high mAP@50 of 0.982 for detection and 0.981 for instance segmentation on the ensemble model, MOTP 0.916 in the KBTrack multi object tracking algorithm and counting accuracy of 0.988 with an RMSE of 0.669. This confirms its ability to accurately detect and track plants. This solution provides a robust framework for addressing the limitations of current methods, offering an effective and scalable approach to modernize ornamental plant inventory systems.