煤矿井下视觉同步定位与地图构建SLAM(Simultaneous Localization and Mapping)应用中,光照变化与低纹理场景严重影响特征点的提取和匹配结果,导致位姿估计失败,影响定位精度。提出一种基于改进定向快速旋转二值描述符ORB(Oriented Fast...煤矿井下视觉同步定位与地图构建SLAM(Simultaneous Localization and Mapping)应用中,光照变化与低纹理场景严重影响特征点的提取和匹配结果,导致位姿估计失败,影响定位精度。提出一种基于改进定向快速旋转二值描述符ORB(Oriented Fast and Rotated Brief)-SLAM3算法的煤矿井下移动机器人双目视觉定位算法SL-SLAM。针对光照变化场景,在前端使用光照稳定性的Super-Point特征点提取网络替换原始ORB特征点提取算法,并提出一种特征点网格限定法,有效剔除无效特征点区域,增加位姿估计稳定性。针对低纹理场景,在前端引入稳定的线段检测器LSD(Line Segment Detector)线特征提取算法,并提出一种点线联合算法,按照特征点网格对线特征进行分组,根据特征点的匹配结果进行线特征匹配,降低线特征匹配复杂度,节约位姿估计时间。构建了点特征和线特征的重投影误差模型,在线特征残差模型中添加角度约束,通过点特征和线特征的位姿增量雅可比矩阵建立点线特征重投影误差统一成本函数。局部建图线程使用ORB-SLAM3经典的局部优化方法调整点、线特征和关键帧位姿,并在后端线程中进行回环修正、子图融合和全局捆绑调整BA(Bundle Adjustment)。在EuRoC数据集上的试验结果表明,SL-SLAM的绝对位姿误差APE(Absolute Pose Error)指标优于其他对比算法,并取得了与真值最接近的轨迹预测结果:均方根误差相较于ORB-SLAM3降低了17.3%。在煤矿井下模拟场景中的试验结果表明,SL-SLAM能适应光照变化和低纹理场景,可以满足煤矿井下移动机器人的定位精度和稳定性要求。展开更多
图像特征提取匹配做为视觉SLAM(Simultaneous Localization and Mapping)的重要组成部分,在井下无人巡检机器人上应用广泛。针对井下环境复杂,光照不足,现有特征提取匹配算法存在匹配率低,进而导致视觉SLAM定位精度低的问题。通过对现有...图像特征提取匹配做为视觉SLAM(Simultaneous Localization and Mapping)的重要组成部分,在井下无人巡检机器人上应用广泛。针对井下环境复杂,光照不足,现有特征提取匹配算法存在匹配率低,进而导致视觉SLAM定位精度低的问题。通过对现有LSD(Line Segment Detector)线特征匹配算法进行改进,采用对比度亮度和对数变换算法对采集的视频图像帧进行图像增强,利用Canny边缘提取算法对增强后的视频图像帧进行图像边缘信息提取后进行LSD线特征提取匹配,与原始算法进行平均匹配率对比分析。结果表明:在连续300帧井下视频图像匹配过程中,改进算法的平均匹配率为99.88%,原始算法的平均匹配率为88.42%,其平均匹配率提升11.46%。说明改进的LSD井下视频图像线特征提取匹配算法具有更高的匹配精度且更适用与井下无人巡检机器人进行无人巡检工作。展开更多
In order to make the humanoid robot walk freely in complicated circumstance, the reliable capabilities for obtaining plane information from its surroundings are demanded. A system for extracting planes from data taken...In order to make the humanoid robot walk freely in complicated circumstance, the reliable capabilities for obtaining plane information from its surroundings are demanded. A system for extracting planes from data taken by stereo vision was presented, After the depth image was obtained, the pixels of each line were scanned and split into straight line segments. The neighbouring relation of line segments was kept in link structure. The groups of three line segments were selected as seed regions. A queue was maintained for storing seed regions, and then the plane region was expanded around the seed region. The process of region growing continued until the queue of seed regions was empty. After trimming, the edges of the planes became smooth. In the end, extracted planes were obtained. In the experiment, two models were used： pipe and stairs. Two planes in pipe mode/and six planes in stairs model were extracted exactly. The speed and precision of algorithm can satisfy the demands of humanoid robot＇s navigation.展开更多
文摘煤矿井下视觉同步定位与地图构建SLAM(Simultaneous Localization and Mapping)应用中,光照变化与低纹理场景严重影响特征点的提取和匹配结果,导致位姿估计失败,影响定位精度。提出一种基于改进定向快速旋转二值描述符ORB(Oriented Fast and Rotated Brief)-SLAM3算法的煤矿井下移动机器人双目视觉定位算法SL-SLAM。针对光照变化场景,在前端使用光照稳定性的Super-Point特征点提取网络替换原始ORB特征点提取算法,并提出一种特征点网格限定法,有效剔除无效特征点区域,增加位姿估计稳定性。针对低纹理场景,在前端引入稳定的线段检测器LSD(Line Segment Detector)线特征提取算法,并提出一种点线联合算法,按照特征点网格对线特征进行分组,根据特征点的匹配结果进行线特征匹配,降低线特征匹配复杂度,节约位姿估计时间。构建了点特征和线特征的重投影误差模型,在线特征残差模型中添加角度约束,通过点特征和线特征的位姿增量雅可比矩阵建立点线特征重投影误差统一成本函数。局部建图线程使用ORB-SLAM3经典的局部优化方法调整点、线特征和关键帧位姿,并在后端线程中进行回环修正、子图融合和全局捆绑调整BA(Bundle Adjustment)。在EuRoC数据集上的试验结果表明,SL-SLAM的绝对位姿误差APE(Absolute Pose Error)指标优于其他对比算法,并取得了与真值最接近的轨迹预测结果:均方根误差相较于ORB-SLAM3降低了17.3%。在煤矿井下模拟场景中的试验结果表明,SL-SLAM能适应光照变化和低纹理场景,可以满足煤矿井下移动机器人的定位精度和稳定性要求。
基金Project(60776816) supported by the National Natural Science Foundation of China and Civil Aviation Administration of ChinaProject(8251064101000005) supported by the Natural Science Foundation of Guangdong Province,China
文摘In order to make the humanoid robot walk freely in complicated circumstance, the reliable capabilities for obtaining plane information from its surroundings are demanded. A system for extracting planes from data taken by stereo vision was presented, After the depth image was obtained, the pixels of each line were scanned and split into straight line segments. The neighbouring relation of line segments was kept in link structure. The groups of three line segments were selected as seed regions. A queue was maintained for storing seed regions, and then the plane region was expanded around the seed region. The process of region growing continued until the queue of seed regions was empty. After trimming, the edges of the planes became smooth. In the end, extracted planes were obtained. In the experiment, two models were used： pipe and stairs. Two planes in pipe mode/and six planes in stairs model were extracted exactly. The speed and precision of algorithm can satisfy the demands of humanoid robot＇s navigation.
