Tomato Recognition and Localization Method Based on Improved YOLOv5n-seg Model and Binocular Stereo Vision

Recognition and localization of fruits are key components to achieve automated fruit picking. However, current neural-network-based fruit recognition algorithms have disadvantages such as high complexity. Traditional stereo matching algorithms also have low accuracy. To solve these problems, this study targeting greenhouse tomatoes proposed an algorithm framework based on YOLO-TomatoSeg, a lightweight tomato instance segmentation model improved from YOLOv5n-seg, and an accurate tomato localization approach using RAFT-Stereo disparity estimation and least squares point cloud fitting. First, binocular tomato images were captured using a binocular camera system. The left image was processed by YOLO-TomatoSeg to segment tomato instances and generate masks. Concurrently, RAFT-Stereo estimated image disparity for computing the original depth point cloud. Then, the point cloud was clipped by tomato masks to isolate tomato point clouds, which were further preprocessed. Finally, a least squares sphere fitting method estimated the 3D centroid co-ordinates and radii of tomatoes by fitting the tomato point clouds to spherical models. The experimental results showed that, in the tomato instance segmentation stage, the YOLO-TomatoSeg model replaced the Backbone network of YOLOv5n-seg with the building blocks of ShuffleNetV2 and incorporated an SE attention module, which reduced model complexity while improving model segmentation accuracy. Ultimately, the YOLO-TomatoSeg model achieved an AP of 99.01% with a size of only 2.52 MB, significantly outperforming mainstream instance segmentation models such as Mask R-CNN (98.30% AP) and YOLACT (96.49% AP). The model size was reduced by 68.3% compared to the original YOLOv5n-seg model. In the tomato localization stage, at the range of 280 mm to 480 mm, the average error of the tomato centroid localization was affected by occlusion and sunlight conditions. The maximum average localization error was ±5.0 mm, meeting the localization accuracy requirements of the tomato-picking robots. This study developed a lightweight tomato instance segmentation model and achieved accurate localization of tomato, which can facilitate research, development, and application of fruit-picking robots.