HFN-SLAM:Hybrid Scene Neural Representation SLAM Based on Frame Alignment and Normal Consistency

Recent advancements in SLAM based on neural radiance fields have demonstrated promising performance. However, existing methods still exhibit shortcomings in terms of reconstruction and pose estimation accuracy, particularly in medium-to-large indoor scenes. These limitations stem from inadequate utilization of structural scene information and ineffective constraint handling for cumulative errors between sequences. To address these challenges, we introduce HFN-SLAM, a neurovisual SLAM system designed to achieve real-time, high-fidelity scene reconstruction and robust camera tracking. To attain fine-grained scene reconstruction without compromising real-time performance, we propose a hybrid representation method for scenes. This method integrates high-resolution, dense 3D hash grid features and 2D plane features, enhancing scene reconstruction accuracy while minimizing parameter overhead. To effectively leverage information between input frames and mitigate accumulated sequence errors, we introduce a frame-aligned algorithm. This algorithm globally aligns input sequences by constraining reprojection errors between keyframes. Furthermore, to enhance scene details, we propose a region-aware normal consistency method. This method implements constraints on large planar scenes, facilitating detailed scene reconstruction. Experimental results demonstrate that our method operates at 3-5 Hz on a desktop PC, surpassing existing methods in both scene reconstruction and camera tracking performance.