Most visual Simultaneous Localization and Mapping (SLAM) systems are built on a fragile lie: the "static world assumption." These algorithms assume that every feature they see—a corner of a table, a crack in the pavement—is a permanent, unmoving fixture of the universe.
In the real world, this assumption collapses. Pedestrians walk across the frame, cars drive by, and doors swing open. When a SLAM system treats a moving person as a fixed landmark, it incorrectly attributes the person's motion to the camera itself. This leads to "trajectory drift," where the estimated path of the robot veers off into space, and "ghosting artifacts," where moving objects leave smeared, translucent trails across the 3D map.
The Geometry of Motion Filtering
The first line of defense against dynamic "noise" is distinguishing between the camera's ego-motion and the independent motion of objects. Early robust frameworks, such as the one proposed by Jun Shimamura, Hideki Koike, and Masashi Morimoto in Robust vSLAM for Dynamic Scenes, used Gaussian Mixture Models (GMM) to analyze the angles of optical flow vectors. If most vectors point in one direction (the background), any vector pointing elsewhere is likely a moving object that should be ignored.
However, geometry alone is often insufficient. A parked car is "potentially" moving but currently static; a SLAM system should use it for localization. But if that car starts to drive away, it becomes a liability. Alejandro Fontán and Javier Civera addressed this in DOT: Dynamic Object Tracking for Visual SLAM by combining instance segmentation with multi-view geometry. By generating precise masks, they allow the system to ignore truly dynamic obstacles while retaining useful static data from "movable" objects that happen to be stationary.
Semantic Awareness and Neural Inpainting
As deep learning matured, researchers began using semantic priors to "pre-clean" the scene. Systems like YG-SLAM by Guoming Chu leverage YOLOv8 to identify pedestrians and vehicles instantly, achieving massive reductions in localization error by simply refusing to track them.
But removing an object leaves a "hole" in the map. VTD-SLAM, developed by Mengyuan Chen and colleagues, uses Vision Transformers to segment dynamic elements and then "inpaints" the resulting gaps using static background data from previous frames. This ensures the map remains geometrically consistent even after the moving object has passed.
The Shift to 4D Gaussian Splatting
The most recent frontier in dynamic SLAM moves away from simply filtering motion to modeling it. Traditional dense SLAM relied on rigid voxel grids or point clouds, which struggle to represent change. New methods utilize 3D Gaussian Splatting (3DGS)—a representation that uses flexible, semi-transparent "blobs" to represent the scene.
In Gassidy, Long Wen introduces the concept of "rendering loss flows." By observing how reconstruction errors evolve over time, the system can pinpoint exactly which Gaussians belong to moving objects.
Taking this further, D2GSLAM by Siting Zhu and Hesheng Wang treats the world as a 4D space. Instead of treating a walking person as noise to be deleted, they reconstruct both the static background and the moving foreground simultaneously. This hybrid 3D/4D approach allows for state-of-the-art tracking because the system understands the entire scene, rather than just a filtered subset of it.
Synergistic Reconstruction
Some researchers have found that dynamic objects aren't just obstacles—they can be assets. In SynCHMR, Tuanfeng Yang Wang and Jimei Yang use the known dimensions of the human body as a "metric ruler." By jointly reconstructing human motion and camera trajectories, the system uses the person in the frame to calibrate the scale of the entire 3D scene, turning a traditional source of SLAM failure into a source of geometric truth.
By moving from simple outlier rejection to sophisticated spatio-temporal modeling, these researchers are enabling robots to navigate not just empty labs, but the chaotic, ever-changing environments of the real world.