转运机器人视觉标签原始图像数据

该数据集面向转运机器人视觉定位与轨道对接区域高精度补盲的研究需求建设，核心背景是在低特征工业场景中，激光雷达常因场景纹理单一、缺乏有效特征点而出现定位漂移问题，严重影响轨道对接精度，而视觉标签定位凭借其局部特征识别精度高的优势，可实现对激光雷达定位的高精度辅助补盲，形成优势互补的定位方案。其资源来源于与激光雷达数据集相同的中车唐山机车有限公司车体事业部涂装厂房南区场景，保证了数据采集环境的一致性，便于后续多传感器数据融合研究。采集设备为转运机器人搭载的Intel RealSense双目相机D455，该相机具备出色的RGB图像采集与深度测量能力，支持在工业环境下的稳定工作，可同步输出高分辨率图像与精准深度信息。采集环境经过严格控制，光照强度稳定在500-1500lux范围内，完全覆盖工业车间白天自然光与夜间人工照明的常见光照工况，同时严格避免强光直射镜头或工件遮挡形成的阴影干扰，确保视觉数据的稳定性。为提升视觉定位的适应性，在轨道对接区域及周边人工布置了不同尺寸、不同距离、不同安装角度的ArUco视觉标签，涵盖了实际作业中标签可能出现的多种姿态。产生方法为通过ROS系统搭建多源数据同步采集框架，精准同步采集相机的RGB图像、深度图像及相机内参信息，同时实时关联机器人的运动状态数据。采集过程重点覆盖标签部分遮挡、光照强度动态波动、标签表面轻微污染等实际工业场景中可能出现的干扰工况，共完成25组不同干扰场景下的轨道对接数据采集。主要内容包含视觉标签及周边环境的原始图像数据、完整的相机标定参数文件、各数据维度的时间戳同步信息。数据格式为ROS标准.bag格式，数据量达5.62GB，包含上万帧图像与深度数据。该数据集既有效支撑了轨道对接区域定位误差≤0.6cm的技术指标验证，也为视觉标签识别算法训练、相机畸变校正、图像特征提取与匹配、图像-点云配准等关键技术研发提供了贴近实际作业场景的多样化数据，显著提升了视觉定位算法在复杂工业环境中的鲁棒性与适应性。

This dataset was constructed for the research needs of visual positioning and high-precision blind-spot supplementation in the rail docking area of transfer robots. The core background is that in low-feature industrial scenarios, LiDAR often suffers from positioning drift due to single scene texture and lack of effective feature points, which seriously affects rail docking accuracy. In contrast, visual tag positioning, with its advantage of high local feature recognition accuracy, can achieve high-precision auxiliary blind-spot supplementation for LiDAR positioning, forming a positioning scheme with complementary advantages. Its data source is the same south area of the painting workshop of the Body Division of CRRC Tangshan Co., Ltd. as the LiDAR dataset, which ensures the consistency of the data collection environment and facilitates subsequent multi-sensor data fusion research. The collection equipment is the Intel RealSense D455 binocular camera mounted on the transfer robot. This camera has excellent RGB image collection and depth measurement capabilities, supports stable operation in industrial environments, and can simultaneously output high-resolution images and accurate depth information. The collection environment is strictly controlled, with the illumination intensity stabilized in the range of 500-1500 lux, which fully covers the common illumination conditions of daytime natural light and nighttime artificial lighting in industrial workshops. At the same time, shadow interference caused by direct strong light on the lens or workpiece occlusion is strictly avoided to ensure the stability of visual data. To improve the adaptability of visual positioning, ArUco visual tags of different sizes, distances and installation angles are manually arranged in and around the rail docking area, covering various postures that the tags may have in actual operations. The generation method is to build a multi-source data synchronous collection framework through the ROS system, accurately synchronously collect the camera's RGB images, depth images and camera intrinsic parameter information, and simultaneously associate the robot's motion state data in real time. The collection process focuses on covering interference conditions that may occur in actual industrial scenarios, such as partial occlusion of tags, dynamic fluctuation of illumination intensity, and slight contamination on tag surfaces. A total of 25 sets of rail docking data collection under different interference scenarios have been completed. The main content includes original image data of visual tags and their surrounding environment, complete camera calibration parameter files, and timestamp synchronization information of each data dimension. The data format is ROS standard .bag format, with a total data volume of 5.62 GB, including tens of thousands of frames of image and depth data. This dataset not only effectively supports the verification of technical indicators with a positioning error ≤0.6 cm in the rail docking area, but also provides diversified data close to actual operation scenarios for the research and development of key technologies such as visual tag recognition algorithm training, camera distortion correction, image feature extraction and matching, and image-point cloud registration, significantly improving the robustness and adaptability of visual positioning algorithms in complex industrial environments.

搜集汇总

背景与挑战

背景概述

该数据集专为转运机器人视觉定位与轨道对接高精度补盲研究设计，旨在解决低特征工业场景中激光雷达定位漂移问题。数据采集于真实工业环境，使用双目相机在受控光照条件下捕获包含ArUco视觉标签的原始图像和深度信息，覆盖多种干扰工况，格式为ROS标准.bag，数据量5.62GB。它支持轨道对接精度验证和视觉算法研发，具有高度的实际应用性和研究价值。