DynaMeshSLAM: A Mesh-Based Dynamic Visual SLAMMOT Method

In order to estimate both camera poses and dynamic object poses, the visual SLAMMOT method combines visual Simultaneous Localization and Mapping (SLAM) with Multiple Object Tracking (MOT). Many visual SLAMMOT methods represent dynamic objects as bounding boxes and point cloud clusters, which ignores the geometric properties of the object surfaces that can provide additional constraints. In this letter, we propose DynaMeshSLAM, a visual SLAMMOT method, which represents dynamic objects as mesh models to leverage intrinsic geometric properties. Firstly, DynaMeshSLAM fuses the mesh projection and the optical flow to achieve multi-level object data association. Secondly, a constrained mesh smoothing method is embedded into the visual SLAMMOT framework to adjust dynamic landmarks depending on both the smoothness of object mesh models and the projection error of mesh vertices. Thirdly, a bundle adjustment solution incorporating the deformation graph optimizes the states of dynamic objects, while ensuring the local rigidity of the smoothed mesh models. Experiments on the KITTI-Tracking dataset demonstrate that our method achieves state-of-the-art performance in both object tracking and object pose estimation.