Abstract
To monitor the growth and structural changes of crop organs, dynamic plant phenotyping based on time-series point clouds has become a cutting-edge research topic. However, existing organ tracking methods based on crop time-series point clouds either rely on complete organ instance segmentation results or lack real-time performance in capturing spatiotemporal correlations among organs. To address these limitations, we propose 3D-OGT: a framework capable of performing continuous organ tracking throughout the entire growth sequence with only the minimal segmentation information. The 3D-OGT framework can automatically propagate organ labels from the previous moment's crop point cloud to the subsequent point cloud, while completing organ segmentation and tracking on multiple crop growth sequences. The framework can recognize and track new organs, mature organs, and even suddenly disappeared organs. Experimental results on a spatiotemporal point cloud dataset demonstrate that 3D-OGT achieves satisfactory organ tracking performance, with an average organ tracking accuracy (TrackAcc) reaching 88.10%, which is superior to three other mainstream methods participating in the comparison.