Cabbage Detection and Counting Using YOLOv8-CBAM and Improved BoT-SORT for Greenhouse and Open Field Cultivation in Korea

Accurate plant counting supports stand assessment and crop management, but canopy overlap, camera motion, and temporary occlusion complicate video-based monitoring. This study developed an RGB image- and video-based framework for detecting, tracking, and counting cabbage seedlings and grown plants under greenhouse and open-field conditions. The framework integrated YOLOv8n with convolutional block attention modules (CBAM) and an improved BoT-SORT procedure incorporating trajectory persistence, short-gap relinking, class-consistent identity management, and counting-event filtering. Detector performance was evaluated across five independent training runs, counting performance across nine held-out videos, and tracking performance on four manually annotated videos. YOLOv8n-CBAM achieved mean precision, recall, mAP@0.50, and mAP@0.50:0.95 values of 0.880 ± 0.007, 0.920 ± 0.007, 0.960 ± 0.006, and 0.600 ± 0.008, respectively (mean ± sample standard deviation). The complete framework achieved 99.27% mean counting accuracy, a mean absolute error of 0.56 plants, a root mean square error of 0.88 plants, and an R2 of 0.997. On the annotated subset, Improved BoT-SORT increased mean IDF1 from 79.6% to 82.3% and mean HOTA from 74.3% to 83.2%, while reducing total identity switches from 23 to 12. Complete-pipeline throughput was 23.38 frames per second on an RTX 3090 workstation. These findings support cabbage population monitoring under the evaluated conditions; independent-site validation, tracking-component ablation, and deployment-device testing remain necessary to establish broader applicability.