A Fruit Detection Algorithm for a Plum Harvesting Robot Based on Improved YOLOv7

The paper presents a novel fruit detection algorithm for a plum harvesting robot. At present, the adequate recognition of plum fruits remains a particularly challenging, under-researched task. Difficulties occur due to small plum fruit sizes and dense growth, as well as numerous occlusions in their environment. A harvesting robot operating in such conditions needs to understand which fruits are reachable, in order to avoid collision and end effector damage. This makes a precise and robust visual detection system of crucial importance. Therefore, a lightweight plum detection procedure based on the improved YOLOv7 algorithm has been proposed. Firstly, the images of domestic plums (Prunus domestica L.) were collected in the field, and train, validation and test sets were established. Secondly, light data augmentation was performed. Next, the initial anchor box sizes used by the original YOLOv7 have been updated, based on the plum sizes in the collected dataset. Finally, an SE (Squeeze-and-Excitation) module was added to the backbone network, which helps model the channel interdependencies at almost no computational cost. The Improved-YOLOv7 model was then trained and evaluated on our dataset. The achieved Precision, Recall and mAP were 70.2%, 72.1% and 76.8%, respectively. The model has been compared with other recent models from the YOLO family, and has shown the best accuracy and generalization ability in real, complex environments. Therefore, the proposed plum detection method can provide theoretical and technical support for harvesting robots in real environments.