YOLOv3-Litchi Detection Method of Densely Distributed Litchi in Large Vision Scenes

Accurate and reliable fruit detection in the orchard environment is an important step for yield estimation robotic harvesting. However, existing methods often target large relatively sparse fruits, but they cannot provide a good solution small densely distributed fruits. This paper proposes YOLOv3-Litchi model based on YOLOv3 to detect litchi fruits visual scenes. We adjusted prediction scale reduced network layer improve ability of dense ensure speed. From flowering 50 days after maturity, we collected total 266 images, including 16,000 then used them construct dataset. Then, k-means++ algorithm cluster bounding boxes labeled data determine priori box size suitable detection. trained improved model, tested its performance, compared with YOLOv2, YOLOv3, Faster R-CNN actual effect F1 value average time as assessed value. The test results show that higher than YOLOv2 0.1, 0.08, 0.05; 29.44 ms faster algorithm, 19.56 ms, 607.06 algorithm. And speed faster. proposed remits optimal performance work presented here may reference further study fruit-detection natural environments.