DAP-VINS: Monocular Visual-Inertial SLAM for Dynamic Environments With Instance Association and Moving Probability Propagation

Visual-inertial simultaneous localization and mapping (VISLAM), which is widely used in autonomous driving, can estimate the ego-motion of vehicles and reconstruct the static environment. Limited by the static-environment assumption, the performance of VISLAM in complex dynamic environments degrades significantly due to the outliers introduced by dynamic objects. To address this issue, a novel monocular visual-inertial simultaneous localization and mapping system called DAP-VINS is proposed to improve the pose accuracy in dynamic environments. In the proposed system, an instance association feature-based module (IAFM) is proposed based on the improved Kuhn-Munkres (KM) algorithm to establish temporal correlations between instance objects in consecutive frames. Based on the probabilistic superposition strategy, an instance ID probability map (IDPM) is constructed to associate instances and landmarks. Dynamic feature detection is achieved with a moving probability propagation model based on a binary Bayes filter (BF-MPP), which can effectively combine geometric constraints and instance information from deep learning. The proposed algorithm is evaluated with public data sets and real-world scenarios. Experimental results demonstrate that DAP-VINS outperforms existing state-of-the-art VISLAM implementations on highly dynamic data sets. Specifically, compared to the results of VINS-Mono, the root mean square error of the absolute trajectory error is reduced by 30%–90% in dynamic scenarios.