Efficient Removal of Multi-Path Effects in Underwater Sonar Imaging for Enhanced 3D Reconstruction

Imaging sonars are essential for underwater robotic perception, providing imagery at extended distances beyond the capabilities of standard optical cameras. Despite the success in 3D reconstruction using the Neural Radiation Field (NeRF) method, the challenge of reconstructing accurate and realistic 3D structures for algorithm development persists, primarily due to noise, especially from multi-path reflections off objects or structures. Based on the principles underlying the multi-path effects of sonar, we propose a probability model to eliminate noise points caused by multi-path effects. Subsequently, we utilize 2D sonar images as supervision and employ the NeRF method to reconstruct more accurate 3D structures. Experimental results on synthetic datasets demonstrate the superiority of our approach.