An optimal alignment framework based on RGBD sensor error characteristics

Abstract

Consumer-grade RGBD sensors that provide both colour and depth information have many potential applications, such as robotics control, localization, and mapping, due to their low cost and simple operation. However, RGBD sensor-based SLAM is still inadequate for many high-precision applications, such as 3D reconstruction and precision localization, due to RGBD sensor's inaccurate depth measurement and easy loss track problem in the less textured environment. Aiming at improving the performance of RGBD based SLAM system, this study focuses on the key component in the SLAM system, the alignment method. Firstly, a comprehensive investigation of the measurement errors of RGBD sensor was conducted. Disparity measurement that depth measurement is calculated from was clarified as the basic measurement of RGBD sensor according to the principles of Structure Light Pattern ranging. The four components of RGBD sensor error: 1) Random Noise, 2) Systematic Bias, 3) Depth Resolution Error, and 4) Image Resolution Error was identified in this study. A novel non-centrosymmetric calibration model was proposed to calibrate the systematic bias and can improve the ranging accuracy by 65% compared with existing centrosymmetric lookup table based calibration model. Experiments and normality tests proved that the random disparity noise of the used RGBD sensor follows a zero-mean normal distribution with a 0.6 standard deviation. The variances of depth measurement and 3D point cloud were further derived from disparity variance and error propagation law for a comprehensive understanding of RGBD error. Secondly, different feature based objective functions used in the alignment method were further optimized according to the discovered error characteristics. The error propagation of the variances of different feature based measurement metrics, including point features, line features, and plane features were thoroughly investigated and rigorously derived in this study. These variances were further used to properly define the weight matrix for optimal alignment result. Experiments results in both public TUM dataset and our experimental dataset have demonstrated the improvements made by the proposed weighted objective functions over traditional unweighted or empirically weighted objective functions. Finally, a new sequential model was proposed in this study to utilize all geometry features. The proposed model is an optimization procedure for multi-feature measurement integration that adaptively selects the suitable objective functions according to an update rule based on DOP value. Compared with the existing arbitrary or empirical weight based multi-feature integrated method, the proposed method can achieve optimal result due to its rigorously defined weight based on disparity and error propagation law. Theoretical assessments demonstrated clear improvements made by integration of different features in both DOP value and system reliability, which indicates the integration of different features can make estimated pose less sensitive to measurement error and easy to detect the measurement outliers. Experiment results show that the proposed method can improve RGB based SLAM in modelling accuracy and alignment capability, especially in challenging less textured scene. Compared with traditional solely feature based alignment method, the proposed multiple feature based method improved lost tracking performance over the state-of-art method by 17.3% and 60% in both public and author collected dataset. 54% and 12.5% improvements in modelling performance over solely feature based alignment method were recorded by the proposed method in typical indoor corridor and challenging tunnel scene.