Answers are grounded in this paper and cite their sources.
In complex agricultural environments, rapid and accurate perception and planning are essential for selective tomato harvesting. Yet many existing approaches remain inefficient and insufficiently accurate, limiting practical deployment. We present a Ghost-CA-enhanced YOLOv5 for target recognition and localization, coupled with a K-G picking sequence planner. The YOLOv5 backbone is replaced with four lightweight downsampling stages built from GhostConv and GhostBottleneck blocks to reduce parameters, and Coordinate Attention (CA) modules are inserted after the multi-scale feature maps to improve detection under occlusion. For bounding-box regression, we adopt the EIoU loss to enhance localization precision. For planning, K-means clusters the picking tasks, and a Gaussian kernel determines both the inter-cluster order and the within-cluster fruit sequence. Experiments show that the detector is 8.9 MB, runs at 25.41 FPS, and achieves 95.12% average detection accuracy. Compared with a sequential strategy and Ant Colony Optimization (ACO), the proposed K-G method yields the shortest traversal path. In continuous multi-target tests, recognition, localization, and picking success rates reach 97.14%, 94.29%, and 91.43%, respectively, with an average picking time of ∼6.7 s per tomato-meeting the operational requirements of tomato-picking robots.
No takes yet. Share an insight, caveat, or question.
Wang et al. (2025) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: