高架草莓果实及采摘点检测算法研究

In complex trellis environments，to improve the accuracy and speed of strawberry harvesting robots in simultaneously detecting fruits and their picking points，a joint detection algorithm for elevated strawberries' ripe fruits and picking points is proposed.Based on the YOLOv8n-Pose benchmark model，a YOLOv8-SY model is introduced.By incorporating a multi‑scale convolution module（MSGConv）and constructing MSGBottleneck and MSGC2f modules，the backbone network is optimized to enhance feature extraction capability.A multi‑level feature fusion module（SDI）is integrated with the neck network of ASF-YOLO to strengthen feature fusion and improve detection accuracy. Combined with fruit images and an auxiliary point compensation localization method，joint detection of ripe fruits and picking points is performed for three types of elevated strawberries.Results show that the improved model achieves P，R，mAP_@0.5，and AP_@OKS=0.5 scores of 95.2%，93.8%，97.4%，and 95.5% in target and keypoint detection，representing improvements of 2.5，2.9，2.6，and 1.7 percentage points over the original model，respectively.In tests on 300 complex real‑world images with various occlusions，the mean errors in the X-axis，Y-axis，and Euclidean distance are -0.65，-0.11，and 5.58 pixels，respectively，with an 86.44% success rate for localization within 10 pixels. This study provides technical support for the development and application of automated harvesting machinery for elevated strawberries.