Dynamic Visual SLAM in Complex Environments

Dynamic Visual SLAM in complex environments addresses the simultaneous localisation and mapping challenge in the presence of moving objects, scene clutter and variable lighting. Traditional frameworks base their estimation on a rigid-scene assumption and extract sparse point features to recover camera pose and construct a global map. In highly dynamic or cluttered settings, such methods suffer from feature drift, false correspondences and incomplete reconstructions. To overcome these limitations, recent systems integrate semantic segmentation models, optical flow analysis and multi-view geometry to identify and remove dynamic features before pose optimisation. Emerging architectures exploit parallel processing threads to decouple the tasks of tracking, segmentation and mapping, thereby preserving real-time performance even when using computationally intensive deep neural networks. In addition, hybrid approaches fuse data from depth sensors, inertial measurements or wheel encoders to reinforce robustness. Some frameworks further employ background inpainting to recover static scene content occluded by dynamic objects, enabling the generation of clean 3D models. These advances have led to accurate six-degree-of-freedom pose estimation and dense or semi-dense reconstructions in indoor and outdoor environments, facilitating applications in autonomous navigation, augmented reality and service robotics, where interactions with moving agents and complex infrastructure are commonplace.

Research from Nature Portfolio

No recent Nature Portfolio content available.

Dynamic Visual SLAM in Complex Environments publication trend

The graph below shows the total number of articles in dynamic visual slam in complex environments across all publications each year (not limited to Nature Index journals).

Technical terms

Simultaneous Localisation and Mapping (SLAM): The process of estimating a sensor platform’s trajectory while incrementally building a map of the environment.

Semantic Segmentation: A pixel-wise classification process using deep neural networks to label objects and regions in an image.

Optical Flow: The apparent motion of image pixels between consecutive frames, used to infer dynamic scene elements.

RGB-D Sensor: A camera that captures both colour (RGB) information and per-pixel depth measurements.

Six-Degree-of-Freedom (6DOF): A description of motion encompassing three translational and three rotational axes.

References

1. RDS-SLAM: Real-Time Dynamic SLAM Using Semantic Segmentation Methods. IEEE Access (2021).
2. A New RGB-D SLAM Method with Moving Object Detection for Dynamic Indoor Scenes. Remote Sensing (2019).
3. RDMO-SLAM: Real-Time Visual SLAM for Dynamic Environments Using Semantic Label Prediction With Optical Flow. IEEE Access (2021).
4. DRE-SLAM: Dynamic RGB-D Encoder SLAM for a Differential-Drive Robot. Remote Sensing (2019).