Optimization for RGB-D SLAM based on plane geometrical constraint

We present an indoor RGB-D SLAM optimization algorithm capable of reducing poses drift based on plane geometrical constraints and reconstructing plane structural model incrementally. Our approach extracts planes from keyframes in backend, merges over-segmented planes, establishes observation constraints between keyframes and global landmark planes, and optimizes poses of keyframes and global landmark planes in a general framework for graph optimization (g2o). Moreover, in order to prevent structural constraints between global landmark planes from being destroyed in optimization process, plane angle structural constraints between global landmark planes observed by the same keyframe are added into optimization graph. We test our optimization algorithm on standard RGB-D benchmarks containing rich plane features, demonstrating that our approach can reduce poses drift and the reconstructed plane structural model covers the most part of planar regions of environment. Furthermore, the application feasibility of augmented reality (AR) is tested using reconstructed plane structural model, demonstrating that plane structural model reconstructed by our approach is suitable for AR application.