基于语义信息与动态特征点剔除的SLAM算法

SLAM algorithm based on semantic information and the elimination of dynamic feature points

传统的同时定位与地图构建(Simultaneous localization and mapping,SLAM)算法在现实场景中易受动态物体及背景的影响,针对该问题提出了一种将语义分割与动态特征点剔除相结合的动态SLAM算法,以实现动态场景地图的构建。首先,根据多层通道注意力和空间注意力机制,构造特征融合网络MulAttenNet(Multilayer attention network),并进行语义分割,剔除场景中运动概率大的物体,粗略估计相机位姿;其次,根据相机位姿和深度信息剔除动态区域;最后,利用剔除后的特征点进行地图的构建。对MulAttenNet网络和动态SLAM算法进行实验,以验证算法的有效性,实验结果表明:该算法构造的MulAttenNet网络能有效提高语义分割的准确性,平均像素准确度提高4.05%,均交并比提高2.60%;将该算法构建的动态SLAM算法与现有SLAM算法相比,建图的绝对位姿误差和相对位姿误差都有所缩小。该算法能在动态场景下构建高精度的语义地图。

Traditional simultaneous localization and mapping(SLAM)algorithms are easily influenced by dynamic objects and their backgrounds in real scenes.In this paper,a dynamic SLAM algorithm that combines semantic segmentation techniques with dynamic feature point rejection has been proposed to establish dynamic scene maps.Firstly,the feature fusion network MulAttenNet based on multi-channel attention and spatial attention was built to perform semantic segmentation by using RGB information,which eliminated objects with high motion probability in the scene and roughly estimated camera poses.Secondly,dynamic areas were eliminated according to camera poses and depth information.Finally,the map was established according to the static feature points.In order to verify the effectiveness of the proposed algorithm in this paper,validation experiments were conducted for the established MulAttenNet network and dynamic SLAM algorithm respectively.The experimental results have shown that the network built in this paper could effectively improve the accuracy of semantic segmentation,the MPA value was increased by 4.05%and the MIoU value was increased by 2.60%.In addition,compared with the existing SLAM algorithms,the dynamic SLAM algorithm established by this algorithm reduced the absolute pose errors and the relative pose errors.The algorithm in this paper can construct highly accurate semantic maps in dynamic scenes.