Abstract: Human-robot collaboration is a promising means to promote orchard intelligence and reduce the over-reliance on manual work for complex agronomic practices such as fruit tree pruning, flower and fruit thinning, and harvesting. Accurate target detection and recognition of robots on humans are the basis and prerequisite for subsequent autonomous human-robot collaboration. In this study, detection and recognition of following robots for human torso were carried out in a standardized hilly orchard. A LiDAR-based human torso detection method was proposed based on the actual orchard environment. Breakpoint detection was used to cluster and segment the point clouds, and the segmentation thresholds were determined based on experimental results. The geometric attributes of the human torso were trained in the classification detection model, resulting in the extraction of six geometric attributes of the human torso. The classification model was then trained with various combinations to obtain the optimal feature combination girth-depth-average curvature (G-D-k) for human torso recognition in an orchard environment. Practical experiments were carried out to validate the feasibility and accuracy of the G-D-k feature combination. The experimental results demonstrate that the G-D-k feature combination can accurately recognize human bodies in orchards. The LiDAR-based detection method can achieve relatively accurate human detection and recognition in complex orchard environments, providing a reference for target detection in human-robot collaboration in orchards.