HMC-SLAM: A Robust VSLAM Based on RGB-D Camera in Dynamic Environment Combined Hierarchical Multidimensional Clustering Algorithm

Most of the current visual simultaneous localization and mapping (SLAM) systems assume that the working environment is static, but dynamic targets in the actual working environment will reduce the position estimation accuracy. Aiming at the problem that visual SLAM cannot filter dynamic targets in a dynamic environment, this article proposes a visual SLAM (VSLAM) system for a dynamic environment based on hierarchical multidimensional clustering (HMC) of feature points, called HMC-SLAM, which can effectively reduce the localization error in dynamic environments. The system identifies a dynamic detection box by combining semantic information and polar geometric constraints. In the dynamic detection box, feature point clusters are divided by fusing the polar distance and depth information of feature points. During the process of dynamic feature point cluster removal, the strongest dynamic point cluster is first identified, and then the adjacent point clusters in its depth range are removed together. To further refine the rejection, we define optimized weights for the feature points in the detection box to avoid the influence of potential dynamic points. In the experiments, we evaluate the performance of the proposed HMC-SLAM in the TUM and BONN datasets and real scenarios. In the TUM and BONN datasets, the algorithmic accuracy and stability of HMC-SLAM are improved by 80.37% and 87.48% relative to ORBSLAM3, respectively. The real scenario results show that HMC-SLAM maintains good localization accuracy, effectively rejects dynamic targets, and is more robust than the current state-of-the-art dynamic SLAM methods.