关键词: 葡萄, 机器人, YOLO-v4, 注意力机制, 目标检测

Abstract:  Aiming at the difficulty for the grape robot to formulate a collision-free picking strategy based on the visual detection results due to the complex and changeable grape growth scene in the orchard environment, a grape detection method based on improved YOLO-v4 in different occlusion states was proposed. First, according to the state of the grape growth scene in the orchard environment, the grapes were marked into four types: no shading, leaf shading, branch shading, and overlapping shading. Then the YOLO-v4 framework was used as the detection model, in which the attention mechanism model (CBAM) was respectively embedded in the backbone network (CSPDarknet53, YOLO-C-C) and path aggregation network (PANet, YOLO-C-P). The network was enhanced by performing target attention on the CSPDarknet53 and PANet network feature extraction processes. Furthermore, the ability to extract grape features reduced the interference of complex scenes in order to achieve high-precision detection of grapes under different occlusions in orchard environments. Finally, the best grape detection model YOLO-C-P for the orchard obscuration scenario was derived by comparing the recognition accuracy and F1 scores of YOLO-C-C and YOLO-C-P networks. The performance evaluation of the method and comparison with other algorithms showed that the accuracy of the YOLO-C-P model was 91.26%, 92.47%, 92.41%, and 90.65% for the grape detection of no shading, leaf shading, branch shading, and overlapping shading, respectively, with the average F1 score of 91.71%. Compared with the same series of models YOLO-v4, YOLO-X-X, YOLO-v5-X, the F1 score increased by 12.62, 8.65, and 5.31 percentage points, respectively. The average time to recognize an image was 0.13s. The research can quickly and effectively identify grapes of no shade, leaf shade, branch shade, and overlapping shade, and can help robots formulate picking strategies (picking order and path planning) in orchard environments so as to avoid collisions caused by occlusions, and consequently provides grape robots with an auxiliary decision-making method for orchard picking.

Key words: grape, robot, YOLO-v4, attention mechanism, object detection