Deep Learning and Stereo Matching Based System for Identification and Localization of Camellia Oleifera Fruits

In order to improve the accuracy of identifying and locating Camellia oleifera fruit among branches, leaves, trunks, fruits, and other complex environments, a system with a visualized interface was designed. It integrates deep learning and stereo matching algorithms. Collecting 2360 pictures construct the dataset, selecting YOLOv5-lite as the recognition algorithm extracting target features. The precision rate of the optimal training model was 93.8%. Compared with the mainstream object detection algorithms YOLOv4-tiny, Faster R-CNN, and SSD, its map is 4.8% greater than other methods and requires just 1-7% of the memory. To achieve more reliable localization results in the real picking environment, the SGBM stereo matching algorithm was introduced to calculate the parallax of each pixel point to estimate the depth. The Euclidean distance between the left camera and the target position is taken as the true depth value. Experiments show that the error between the estimated depth and the true depth can be controlled within 1 cm. Furthermore, this system combines hardware and software, detecting fruits quickly in dark, exposed, occluded, and dense conditions. Providing visual support for automated picking systems while also encouraging the integration of computer vision, image processing, control, and other fields.