Detection of multi-class coconut clusters for robotic picking under occlusion conditions

With the development of tree-climbing robots and robotic end-effectors, it is possible to develop automated coconut-picking robots with the help of machine vision technology. Coconuts grow in clusters in the canopy and are easily occluded by leaves. Therefore, the detection of multi-class coconut clusters according to the occlusion condition is necessary for robots to develop picking strategies. The coconut detection model, named YOLO-Coco, was developed based on the YOLOv7-tiny network. It detected coconuts in different conditions such as not-occluded, leaves-occluded, and trunk-occluded fruit. The developed model used Efficient Channel Attention (ECA) to enhance the feature weights extracted by the backbone network. Re-parameterization Convolution (RepConv) made the model convolution layers deeper and provided more semantic information for the detection head. Finally, the Bi-directional Feature Pyramid Network (BiFPN) was used to optimize the head network structure of YOLO-Coco to achieve the balanced fusion of multi-scale features. The results showed that the mean average precision (mAP) of YOLO-Coco for detecting multi-class coconut clusters was 93.6%, and the average precision (AP) of not-occluded, leaves-occluded, and trunk-occluded fruit were 90.5%, 93.8%, and 96.4%, respectively. The detection accuracy of YOLO-Coco for yellow coconuts was 5.1% higher than that for green coconuts. Compared with seven mainstream deep learning networks, YOLO-Coco achieved the highest detection accuracy in detecting multi-class coconut clusters, while maintaining advantages in detection speed and model size. The developed model can accurately detect coconuts in complex canopy environments, providing technical support for the visual system of coconut-picking robots.