Time-of-flight-based advanced surface reconstruction methods for real-time volume estimation of bulk harvested wild blueberries

• The voxel grid method emerged as the most efficient and accurate for volume estimation, achieving the lowest mean absolute error-bias corrected (MAEbc) value of 0.0119 m³ and the fastest processing time of 0.012 seconds. • The alpha shape method exhibited substantial error reduction post-bias correction but was significantly hindered by its long processing time, averaging 49.265 seconds. • The octree method demonstrated minimal error reduction post-bias correction, compromising precision compared to the other approaches. • YOLOv8n-seg model effectively enhanced the preprocessing of point cloud data by eliminating outlier points and refining data for volume estimation. • The Basler Blaze-101 3D camera proved effective for data acquisition under varying lighting conditions and for generating accurate point clouds for volume estimation. This study focuses on volume estimation of mechanically harvested wild blueberries contained within a harvester tote, an economically significant agricultural product for northeastern North America. Accurate volume estimation of mechanically harvested berries in their respective totes is essential for improving the efficiency of the harvesting operation. This study evaluated four computational methods for volume estimation including convex hull, alpha shape, octree, and voxel grid. The evaluation of each method was based on three key metrics: Mean absolute error (MAE), Mean absolute error bias corrected (MAEbc), and processing time. Convex hull exhibited the highest initial MAE (0.0495 ± 0.0235 m³) but showed significant error reduction post-bias correction (to 0.0122 ± 0.0128 m³), leading to a reduced variability. Alpha shape resulted in a moderate MAE (0.0379 ± 0.0217 m³) and bias-corrected error (0.0190 ± 0.0134 m 3 ), but its utility was diminished by an extended processing time (49.265 ± 2.841 s). Octree had a lower initial MAE (0.0447 ± 0.0203 m³), with large variability in bias correction effectiveness (0.0394 ± 0.0233 m³) and a moderate processing time (0.266 ± 0.015 s). Voxel grid had superior accuracy and efficiency, presenting the lowest MAE (0.0311 ± 0.0349 m³), MAEbc (0.0119 ± 0.0158 m³), as well as processing time (0.012 ± 0.001 s). These findings emphasize the importance of considering average performance and variability in method selection, particularly in fields where consistent performance is as crucial as efficiency.