摘要
针对传统同时定位与地图构建(SLAM)算法在动态遮挡场景下存在的被遮挡物与动态目标边缘区分模糊、动态特征点误判导致静态特征点不足及跟踪丢失问题,提出一种基于改进YOLOv6s网络的SLAM算法。首先,通过YOLOv6s网络跟踪动态目标并提取语义信息,实现动态目标边界的精准识别与分割。其次,系统同时提取点特征与线特征,并结合改进的Z-score评分机制与对极几何约束筛除动态特征点,保留高质量静态特征点。然后,基于双向映射背景修复模型填补被剔除区域的静态信息,以增强关键帧特征完整性。最后,在TUM公开数据集上进行实验。实验结果表明,相较于DynaSLAM算法,所提算法在动态遮挡场景下的平均绝对轨迹误差降低14.29%,在地图构建质量与轨迹精度方面表现优越。
To address the challenges faced by traditional simultaneous localization and mapping(SLAM)algorithms in dynamic occlusion scenarios—specifically,their difficulty in accurately distinguishing the edges of occluded objects and dynamic targets,leading to segmentation ambiguity,and the misclassification of dynamic feature points,which results in an insufficient number of static feature points and subsequent tracking failures—a SLAM algorithm based on an improved YOLOv6s network is proposed.Firstly,the YOLOv6s network is employed to track dynamic targets and extract semantic information,enabling precise identification of dynamic object boundaries and achieving accurate segmentation.Secondly,the system simultaneously extracts point features and line features,and utilizes an improved Z-score scoring method combined with epipolar geometry constraints to filter out dynamic feature points while preserving high-quality static feature points.Then,a bi-directional mapping-based background restoration model is applied to fill in the removed regions with static information,thereby enhancing the feature integrity of SLAM keyframes.Finally,experiments are conducted on the TUM public dataset.Experimental results show that compared to DynaSLAM algorithm,the proposed method reduces average absolute trajectory error by 14.29%in dynamic occlusion scenarios,exhibiting superior performance in both map construction quality and trajectory accuracy.
作者
陈孟元
许瑞珩
杨苏朋
丁帅
CHEN Mengyuan;XU Ruiheng;YANG Supeng;DING Shuai(School of Electrical Engineering,Anhui Polytechnic University,Wuhu 241000,China;Key Laboratory of Advanced Perception and Intelligent Control for High-end Equipment,Ministry of Education,Wuhu 241000,China)
出处
《中国惯性技术学报》
北大核心
2025年第8期802-811,共10页
Journal of Chinese Inertial Technology
基金
安徽省重点研究与开发计划项目(202304a05020073)
安徽省高校协同创新项目(GXXT-2021-050)
安徽省高校杰出青年科研项目(2022AH020065)。
关键词
同时定位与地图构建
物体遮挡
目标检测
运动判断
背景修复
simultaneous localization and mapping
object occlusion
target detection
motion judgment
background repair
