Objective The study was conducted to improve the detection accuracy of muskmelon picking robot in greenhouse under complex light changes and branch and leaf occlusion, and realize the spatial coordinate positioning of detection targets.

Method Based on YOLOv3, the study explored the impacts of optimizing the combination of different backbone networks, head and neck network structures and bounding box loss function on the model detection performance, established a target detection network model YOLOResNet70 under severe muskmelon occlusion, and then fused the model with Intel RealSense D435i depth visual sensor for target space positioning.

Result With ResNet70 as the backbone network, YOLOResNet70 had the best performance with the combination of SPP (Spatial pyramid pooling), CIoU (Complete Intersection over Union), FPN (Feature Pyramid Network) and NMS (Greedy non-maximum suppression). The average accuracy (AP) of the model reached 89.4%, which was better than 83.3% of YOLOv3 and 82% of YOLOv5, and the detection speed (61.8 frames/s) was 14% faster than that of YOLOv4 (54.1 frames/s).

Conclusion Through the detection and test of occluded muskmelon images under different lighting conditions, it shows that the YOLOResNet70 model has good robustness, and the model is fused with Intel RealSense D435i depth visual sensor to achieve the spatial positioning coordinates of muskmelon, which is consistent with the manual measurement result. It provides theoretical and model support for target detection and spatial positioning of muskmelon picking robot.