基于运动避免特征提取的动态视觉SLAM方法

Dynamic Visual SLAM Method Based on Motion Avoidance Feature Extraction

SLAM是移动机器人最基本的功能。传统的SLAM方法服从静态世界假设,然而真实世界经常包含行人、车辆等动态物,将其作为参照物将导致错误的定位结果。现有的动态SLAM方法大多基于语义分割检测和剔除动态物,但语义分割模型通常计算量很大,使得该类方法难以满足实时性要求。因此,提出一种高效的动态视觉SLAM方法。基于目标检测和运动一致性验证实现一种运动避免的特征提取方法,然后结合著名的ORB-SLAM2实现动态SLAM方法MA-SLAM。实验结果表明,MA-SLAM能够解决动态物造成的问题,并且相比于基于特征过滤的动态SLAM方法,其能够获得数量更充足的特征,从而获得更高的定位精度和鲁棒性。此外,MA-SLAM追踪一帧仅需50 ms,适用于实时任务。

SLAM is important for mobile robots.Dynamic objects,such as pedestrians and vehicles,lead to a poor localization performance in SLAM.As a resolution,several dynamic SLAM methods were proposed in recent years.Most of the existing dynamic SLAM methods are based on semantic segmentation,in the purpose to detect dynamic objects and hence eliminate them from the map.Semantic segmentation models usually involve a large amount of computation,preventing such methods from real-time applications.An efficient dynamic visual SLAM method was proposed.A motion avoidance feature extraction method was realized based on object detection and moving consistency check.A dynamic SLAM method called MA-SLAM(Motion-Avoidance SLAM)was proposed by incorporating ORB-SLAM2 with the feature extraction method.Experimental results show that MA-SLAM can solve the problem caused by dynamic objects effectively.Compared with filtering-based dynamic SLAM methods,MA-SLAM obtains more features and consequently improves the localization accuracy and robustness.Besides,MA-SLAM takes only 50 ms to track a frame,and hence is suitable for real-time tasks.