UcPIN-SLAM: LiDAR SLAM Using Implicit Neural Representations with Uncertainty Modeling

The incorrect estimation of the signed distance function (SDF) values can have a significant impact on the localization and mapping of implicit neural dense SLAM, which is essential for autoonomous robots. To tackle the aforementioned issue, we present UcPIN-SLAM, a LiDAR SLAM using implicit neural representation with uncertainty modeling, to address the impact of noise variability under different incident angles on SLAM for autonomous robots navigation. Firstly, the noise heteroscedasticity parameters are calculated by the incident angle utilizing the lidar measurement noise model. Secondly, the maximum likelihood estimation is achieved according to the noise heteroscedasticity parameters as well as the Laplace distribution, while utilizing a self-decoder trained online via gradients for SDF value corrections. Thirdly, a multi-layer perceptron (MLP) model with loss function,which consists of endpoint constraints(END), maximum probability density function constraints(UNC), binary cross entropy (BCE) and Eikonal regularisation(ENIKNOL), is applied to further SDF calibration. Finally, the sparse optimizable neural points are defined to initialize and real-time update the implicit neural map, constructing a globally consistent high-fidelity implicit neural grid map. Two publicly available outdoor LiDAR datasets are evaluated separately for localization and mapping performance validation. The experiments demonstrate that UcPIN-SLAM achieves positioning accuracy comparable to state-of-the-art LiDAR odometry or SLAM systems, and its mapping quality outperforms the current best LiDAR SLAM methods by up to 15.7% in key metrics.