摘要
针对同时定位与地图构建(SLAM)中动态物体干扰导致系统精度下降的问题,提出一种基于改进实时目标检测算法(RT-DETR)的视觉SLAM方法。该方法融合RT-DETR的检测结果与光流阈值法剔除动态物体特征点,提升了动态场景下的算法性能。改进后的RT-DETR采用轻量化模块替换原始网络主干,在精度仅下降5%的情况下,参数量相比原模型减少45%,大幅降低计算开销并支持移动端部署。在TUM和BONN动态数据集上的测试结果表明,通过动态点剔除,所提方法在动态环境中的绝对轨迹误差与相对位姿误差的均方根误差平均值相比ORB-SLAM3降低约61.76%和36.87%。
To address the problem of decreased system accuracy caused by the interference of dynamic objects in Simultaneous Localization and Mapping(SLAM),a visual SLAM method based on the improved real-time detection Transformer(RT-DETR)is proposed.This method combines the detection results from RT-DETR with an optical flow threshold method to eliminate feature points of dynamic objects,thereby enhancing performance in dynamic scenes.The improved RT-DETR utilizes lightweight modules to replace the original network backbone,achieving a 45%reduction in the number of parameters compared to the original model,while maintaining only a 5%decrease in accuracy.This significant reduction in computational overhead makes deployment on mobile devices feasible.Testing on the TUM and BONN dynamic datasets shows that,due to the removal of dynamic points,the average root mean square error of absolute trajectory error and relative pose error of the proposed method in dynamic environments is reduced by approximately 61.76%and 36.87%compared with ORB-SLAM3.
作者
孙进
申学
SUN Jin;SHEN Xue(College of Internet of Things,Nanjing University of Posts and Telecommunications,Nanjing 210003,China;State Key Laboratory of Ocean Engineering,Shanghai Jiao Tong University,Shanghai 200240,China)
出处
《中国惯性技术学报》
北大核心
2025年第8期794-801,811,共9页
Journal of Chinese Inertial Technology
基金
国家自然科学基金(62203231)
海洋工程国家重点实验室(上海交通大学)开放基金资助(GKZD010084)
江苏省研究生科研与实践创新计划(KYCX24_1206)。
