Abstract: Aiming at the problem that the recognition technology of harvesting robots in crop picking was limited by complex background interference in unstructured environments, especially due to occlusion by foliage and the overlapping of fruits, resulting in lower accuracy in identification, an improved YOLOv5 algorithm was proposed based on the improved research approach involving post-processing of the model. Initially, the centroid distance of fruit targets, the actual difference in predicted box width and height, and the intersection-over-union of areas were collectively considered as loss terms. This was aimed at enhancing the accuracy of predicted box sizes. Furthermore, the centroid distance was utilized as a penalty term weighted by the intersection-over-union score to improve the recognition capability for densely clustered targets. Subsequently, auxiliary training heads were incorporated to provide additional gradient information, thereby preventing overfitting. Through comparative analysis of loss values using multiple loss functions and assessing the model improve mentaccuracy, the effectiveness of the enhancements was experimentally validated. Finally, the deployment onto the robot confirmed the feasibility of the proposed improvements. The results indicated that the improved algorithm model achieved an average accuracy of 95.6%, with a recall rate of 90.1%. Compared to the pre-improvement overall class accuracy, there was an increase of 0.4 percentage points in both accuracy and recall rate, meeting the recognition requirements for harvesting robots.