SDF SLAM: CNN-Driven Real-Time Mapping of Unknown Environment and Path Planning

The problem of environment mapping with path planning in unknown domain can be resolved by SDF SLAM (Signed Distance Function Simultaneous Localization and Mapping) to allow the Model to map its surroundings in realtime, without any prior map. A Region of Interest (ROI) in the corrected image pair is selected for optimization and generates a depth map. A 3D map is built using the depth data from stereo vision, along with model positioning and camera angle. By combining add on data from the model and camera positioning, reliable path planning is achieved, helping ensure smooth and precise navigation. The processing time of the system is reduced through ROI implementation with path planning a* algorithm. A dataset containing RGBD depth images with camera pose estimates further improves the CNN's ability to perform accurate path planning. The CNN's performance is tested based on training settings, parameter adjustments, and accuracy. The final results show that the algorithm reaches 64% accuracy with a loss of about 0.34. For real-time applications the accuracy can be improved by combining model orientation with camera pose estimates and fine-tuning parameters reports as best with other methods.