顾及动态物体感知的增强型视觉SLAM系统

Enhanced Visual SLAM System Considering Dynamic Objects

传统的同步定位与制图(Simultaneous localization and mapping,SLAM)系统在复杂环境下工作时,无法分辨环境中的物体是否存在运动状态,图像中运动的物体可能导致特征关联错误,引起定位的不准确和地图构建的偏差。为了提高SLAM系统在动态环境下的鲁棒性和可靠性,本文提出了一种顾及动态物体感知的增强型视觉SLAM系统。首先,使用深度学习网络对每一帧图像的动态物体进行初始检测,然后使用多视图几何方法更加精细地判断目标检测无法确定的动态物体区域。通过剔除属于动态物体上的特征跟踪点,提高系统的鲁棒性。本文方法在公共数据集TUM和KITTI上进行了测试,结果表明在动态场景中定位结果的准确度有了明显提升,尤其在高动态序列中相对于原始算法的精度提升在92%以上。与其他顾及动态场景的SLAM系统相比,本文方法在保持精度优势的同时,提高了运行结果的稳定性和时间效率。

When working in complex scenarios,traditional simultaneous localization and mapping(SLAM)systems cannot distinguish whether the visible objects are moving.Moving objects in the images may lead to wrong feature association,resulting in the inaccuracy of positioning and the deviation of mapping.To improve the robustness and reliability of the SLAM system in dynamic scenarios,an enhanced visual SLAM system with dynamic object perception is proposed in this paper.Firstly,the object detector is used to initially detect the dynamic objects in each image,and then the multi-view geometry method is further used to extract the dynamic regions that cannot be determined by the object detection.The robustness of the system is improved by eliminating feature points belonging to dynamic objects.The proposed method is tested in public datasets TUM and KITTI.The results show that the localization accuracy of the proposed method in dynamic scenes has been significantly improved,especially in high dynamic sequences.Compared with the original algorithm,the accuracy is improved by more than 92%.Compared with other SLAM systems in dynamic scenarios,the proposed method not only maintains the accuracy advantage,but also improves the stability of running results and time efficiency.