无人艇在复杂水面环境下的自主导航能力是其完成养殖作业的基础,其在智能水产养殖中的发展前景巨大.视觉同时定位与地图构建(Simultaneous location and mapping,SLAM)技术可以提供实时的环境信息,是实现无人艇自主导航的关键.然而,水面环境是一种缺乏足够有效特征点的低纹理场景,且受水面波纹和反光影响存在大量动态无效特征点,导致视觉SLAM位姿的计算精度较差、性能严重下降.为此,提出了一种面向动态水面环境的基于视觉语义和点线融合的SLAM系统.首先,对ORB-SLAM3算法框架进行改进,增加语义分割线程,利用语义信息生成掩码消除水面无效特征点,以消除动态水面环境的干扰.其次,加入线特征来加强系统稳定性,提出了一种基于几何约束的线段匹配方法,提高水面线特征提取和跟踪的准确性,并利用点线特征融合提高数据关联的准确性,解决传统SLAM算法在水面低纹理场景中提取不足的问题.在USVInland数据集上的实验结果显示,与ORB-SLAM3和PL-SLAM算法相比,改进后算法的定位精度在直线航行中平均提高了 44.74％和55.48％,在机动航行中最多提高了 76.60％和70.15％,有效消除了水面干扰对位姿估计的影响,提升了视觉SLAM系统在水面低纹理场景中位姿估计的准确性和鲁棒性.
Dynamic water surface environment perception SLAM algorithm based on visual semantics and point-line fusion for unmanned surface vessels
Unmanned surface vessels(USVs)have autonomous navigation capacities under complex water surface environ-ment,which is the basis for aquaculture operation,with great development prospects in intelligent aquaculture.Simultaneous localization and mapping(SLAM)technology provides real-time environmental information,playing a pivotal role in enabling autonomous navigation for USVs.However,water surfaces present low-texture scenes with insufficient distinct features,com-pounded by dynamic invalid features caused by surface ripples and reflections,which result in poor calculation accuracy and seri-ous performance degradation of visual SLAM pose.To address this problem,the paper proposes a SLAM system oriented to-wards dynamic water surface environments based on visual semantics and point-line fusion.Firstly,the ORB-SLAM3 frame-work is enhanced by introducing a semantic segmentation thread to eliminate invalid water features using semantic masks,thereby mitigating dynamic water interference.Secondly,the system stability is enhanced by incorporating line features to pro-pose a geometric-constraint-based line-matching method to improve the accuracy of waterline feature extraction and tracking.Moreover,the point-line feature fusion is utilized to enhance data association accuracy,resolving the deficiency of traditional SLAM algorithms in extracting features from low-texture water surfaces.The results of the USVInland dataset demonstrate that compared with ORB-SLAM3 and PL-SLAM algorithms,the positioning accuracy of the improved algorithm has averagely in-creased by average of 44.74％and 55.48％in straight-line navigation,and up to 76.60％and 70.15％in maneuvering navigation,which mitigates the impact of water surface disturbances on pose estimation effectively,and enhances the accuracy and robust-ness of visual SLAM systems under low-texture water environment.
Unmanned surface vehicleIntelligent aquacultureVisual SLAMPoint and line featuresSemantic segmentationMachine vision
万方数据
维普
