DMN-SLAM: Multi-MLPs Neural Implicit Representation SLAM for Dynamic Environments

Addressing the limitations of traditional V-SLAM in large-scale and dynamic environments, this paper proposes a multi-MLPs neural implicit representation SLAM in Dynamic Environments(DMN-SLAM), which leverages neural radiance fields to facilitate local mapping in dynamic scenes. This system aims to achieve high-fidelity 3D scene reconstruction, making it suitable for real-world autonomous driving applications. Specifically, the proposed DMN-SLAM system includes a pre-processing semantic segmentation module that processes the scene to eliminate dynamic objects, thereby reducing their impact on SLAM localization and mapping. In the tracking module, a traditional feature-based visual odometry module is employed to estimate poses, enhancing tracking accuracy. In the mapping module, five different MLP neural implicit encoding structures are implemented to efficiently extract and represent detailed features of the scene, enabling local updates and optimization. Finally, a dynamic and efficient octree structure is introduced to store voxel features. This structure not only significantly reduces storage requirements but also allows for rapid querying and updating of map points, further enhancing the performance and practicality of system. Experimental results demonstrate that the proposed method achieves high reconstruction accuracy and quality without the need for pretraining. DMN-SLAM outperforms other neural radiance field SLAM methods in terms of tracking accuracy and mapping quality in dynamic environments.