TreePillars: An end-to-end 3D target recognition algorithm for nursery and orchard spray robot

The dataset used in this study includes 1000 frames of point cloud data featuring scenes of osmanthus and cherry blossom trees. The data were collected using a Velodyne 16-line LiDAR during three distinct acquisition periods to capture seasonal variations in tree structure: November 2022, May 2023, and April 2024. This approach was intentionally designed to ensure a comprehensive and diverse dataset, with each period reflecting different stages in the trees’ growth cycle. The initial data collection in November 2022, referenced from Liu et al. (2024b,a), provides a baseline to observe changes in tree characteristics over time.
The trained TreePillars algorithm achieves an mAP of 61.34%, a computation complexity of 7.34 GFLOPs, and a parameter count of 4.24 million, respectively. Compared to PointPillars, the average precision is increased by 10.94%, while the computation complexity and parameters count are reduced by 4.6% and 12.4%, respectively. It has a more accurate recognition effect on sparse and occluded point cloud targets. Experiments on the intelligent robot show that the location relative error for trees, mAP, and average time for single-frame detection of TreePillars are 1.17%, 77.44%, and 19.14ms, respectively. Compared to the improved DBSCAN and lightweight PointNet, the location relative error is reduced by 1.88%, with detection results showing precise bounding boxes for targets, and a detection time shortened by 76.77%. Relative to the VoteNet algorithm, TreePillars reduces the positiong error by 0.7%, improves mAP by 26.33%, and shortens detection time by 98.29%, meeting the requirements of real-time target detection.