การออกแบบและพัฒนาระบบมองเห็นสำหรับระบบทำสวนยางพาราอัตโนมัติ

Abstract

This thesis presents a machine vision for an automatic rubber latex-farming robot. The design adopts two RGB-Depth cameras (Intel RealSense D400 Series) with the additional lights that aim under the low-light rubber tree orchards. Two RGB-Depth cameras face a rubber tree at two different depth distances wherefore the tapping-panel coverages and the details of the tapping path. A far-range image dataset is build up with bounding box annotations. The thesis includes two different detection algorithms that handle the RGB and the depth image of the far-range image dataset. A color-based with a sliding window algorithm can detect the tapping position bounding box up to 35.4% average precision at 0.5 loU. A CNN-based detection algorithm, Faster-RCNN with pre-trained MobileNetV2, achieves 80.3% average precision of tapping position and cup detections at 0.5 loU. For a near-range image dataset, the thesis manifests the annotation of the tapping line using a bounding box and a polygonal curve with a refinement algorithm, and a tapping line detection algorithm that benefits the tapping path shadow for extracting the line. The evaluation of the near-range algorithm is in two steps, which are the bounding box to obtain the detection precision and then measures the distance error of the detected results. The algorithm produces 89.9% average precision at 0.5 loU for the bounding box and the average distance error at 13 pixels using Hausdorff Distances within the high-resolution 1280 by 720-pixel images, which equals to 9.0 millimeters respected to camera geometry.