Accurate tomato fruit measurement via parallax-based binocular vision and deep learning detection

ABSTRACT: In robotic harvesting and yield prediction scenarios, real-time visual recognition of tomato fruit size often requires additional distance sensors or reference objects for measuring dimensions. This study introduces a low-cost, single-device, variable-distance vision recognition solution. The measurement system employs a binocular camera, and the algorithm used is You Only Look Once version 5 (YOLOv5). The measurement principle is based on parallax, which calculates the distance between the tomato fruit and the camera baseline by analyzing the offset of the same point on the left and right images. The dataset was trained with YOLOv5, including 100 binocular and 500 monocular images of single, multiple, and shielded fruits. The baseline distance between the tomato fruit and the binocular camera was calculated by measuring the parallax principle of the binocular camera. The detection accuracy of ripe tomatoes reached 86.9 % using the YOLOv5 target detection algorithm and various mixed datasets. The average relative errors for distance, vertical width, and horizontal length were 4.92 %, 5.31 %, and 5.05 %, respectively. The test system costs less than US$30. The results demonstrate that the system exhibits excellent dimensional detection performance at a low cost and across variable distances.