关键词: 智慧农业, 观光农业, 桃子, 采摘识别, 深度学习, 人体姿态

Abstract: For the needs of intelligent management of peach picking tourism orchards, a deep learning-based picking recognition method is proposed. The method uses machine vision and deep learning technologies to integrate a lightweight human posture estimation algorithm Lightweight OpenPose, a target detection algorithm YOLOv5s, and a target tracking algorithm DeepSORT to develop a peach picking behavior detection approach. It can be divided into three steps according to the functional order: the picking posture determination method based on the human body joint angles, the picking target determination method based on nearest neighbor retrieval and its optimization, and the picking target detection failure solution method based on the set status flags. A dataset is established based on the actual peach picking videos for performance tests. Comparing the method based on the angle of human joints proposed in this paper with the traditional method of using bounding boxes enclosing human joints, the method in this paper can improve the precision of determination （P） rate of hand-raising action by 16%. For the problem of determining the picking target, the nearest neighbor retrieval approach outperforms both the traditional method based on the comparison of distance and reference size and the method based on the comparison of IoU and thresholds, with an increased P rate by at least 11%. The picking target detection failure solution method based on set status flags effectively solves the influence of hand occlusion on the detection results, substantially improving the P rate by 39%. On this basis, an experimental system is designed to test the proposed method under real-world conditions. The results show that the proposed peach picking recognition method achieves effective and accurate recognition of picking actions in actual orchard environments.

Key words: smart agriculture, agritourism, peach, picking recognition, deep learning, human postures