RTrack-DL: A dynamic scene-oriented model for instance identification and continuous tracking of rice seedling rows

Visual navigation is critical for intelligent paddy field operations, where continuous and accurate tracking of rice seedling rows is essential. However, existing methods predominantly address static frame identification and lack robust solutions for continuous tracking in dynamic field environments, particularly under challenging conditions like missing seedlings and irregular planting patterns. To overcome these limitations, this study develops RTrack-DL, a novel dynamic scene-oriented model for instance-level identification and continuous tracking of rice seedling rows. The novelty lies in its dynamic framework that integrates a dedicated merge-and-deduplication strategy (IoU-Plus) to enhance line-fitting robustness against missing seedlings and irregular planting, and fuses the fitted lines with a multi-object tracker to maintain stable row identities and prevent track loss during field operations. The model employs DeepLabv3+ for pixel-level semantic segmentation of row centerlines. Subsequently, connected component analysis and line fitting are applied to obtain individual instances. To handle discontinuities caused by missing seedlings and irregular planting, a merge-and-deduplication strategy based on an IoU-Plus metric is introduced to eliminate redundant lines and enhance fitting robustness. A multi-object tracking module combining Kalman filtering and the Hungarian algorithm ensures consistent identity tracking across frames. The performance of RTrack-DL is evaluated through experiments under three illumination conditions—sunny, cloudy, and overcast—while considering scenarios involving missing seedlings and irregular planting. Results show optimal performance under cloudy conditions, with a fitting accuracy of 99.42% and an angular error of 0.49°, while sunny and overcast conditions yielded slightly reduced accuracy. In tracking tests, RTrack-DL achieved multiple object tracking accuracy scores of 93.5%, 92.1%, and 92.7% under cloudy, sunny, and overcast conditions, respectively, with real-time performance maintained at 15-20 FPS. The proposed framework demonstrates a practical solution for continuous tracking of dense rice seedling rows in dynamic field environments, providing a robust visual perception component for intelligent navigation systems in fully unmanned paddy field management.