动态环境下基于动态区域剔除的双目视觉SLAM算法

Binocular Vision SLAM Algorithm Based on Dynamic Region Elimination in Dynamic Environment

现有的同步定位与地图创建(SLAM)算法在动态环境中的定位与建图精度通常会大幅度下降,为此提出了一种基于动态区域剔除的双目视觉SLAM算法.首先,基于立体视觉几何约束方法判别场景中动态的稀疏特征点,接下来根据场景深度和颜色信息进行场景区域分割;然后利用动态点与场景分割结果标记出场景中的动态区域,进而剔除现有双目ORB-SLAM算法中动态区域内的特征点,消除场景中的动态目标对SLAM精度的影响;最后进行实验验证,本文算法在KITTI数据集上的动态区域分割查全率达到92.31%.在室外动态环境下,视觉导盲仪测试中动态区域分割查全率达到93.62%,较改进前的双目ORB-SLAM算法的直线行走定位精度提高82.75%,环境建图效果也明显改善,算法的平均处理速度达到4.6帧/秒.实验结果表明本文算法能够显著提高双目视觉SLAM算法在动态场景中的定位与建图精度,且能够满足视觉导盲的实时性要求.

In dynamic environments, the localization and mapping accuracy of the existing simultaneous localization and mapping(SLAM) algorithms will decrease dramatically. For this problem, a binocular vision SLAM algorithm based on dynamic region elimination is proposed. Firstly, the dynamic sparse feature points in the scene are identified by the stereo vision based geometric constraint method, and the scene area is segmented based on the scene depth and color information.Secondly, the dynamic points and the scene segmentation results are used to mark the dynamic regions in the scene, and then eliminate the feature points in the dynamic regions in the existing binocular ORB-SLAM algorithms as well as the impact of dynamic targets in the scene on SLAM accuracy. Finally, the experimental verification of the proposed algorithm is carried out. The recall rate of dynamic region segmentation on the KITTI dataset can reach 92.31%, and it can reach 93.62% in the test of visual guidance in the outdoor environment. Compared with the previous binocular ORB-SLAM algorithm, the straight walking localization accuracy is improved by 82.75%, and the mapping effect is also enhanced significantly. The average processing rate of the algorithm can reach 4.6 frame/s. The results show that the proposed algorithm can significantly improve the localization and mapping accuracy of the binocular vision SLAM algorithm in dynamic scenes, and meets the real-time performance requirement of visual guidance.