Abstract
The significance of three-dimensional point cloud data in generating high precision point cloud maps for environmental sensing, geo-spatial analysis and autonomous driving is increasing today with the advancements in sensing technologies such as LiDAR. Accurate registration is required to reduce noise and maintain uniform point density. The existing algorithms for this purpose shows limitations such as slow convergence, partial overlaps, and convergence to local minima or maxima as a consequence of the non-homogeneous distribution of feature points in the point cloud scenes. To tackle these problems we present a novel scan matching framework Robust and Adaptive Normal Distribution Transform (RANDT). In our method an incremental scan matching module is introduced which continuously perform scan matching with newly matched scans to achieve uniform and dense point distribution across the scene. Prior to the scan matching process an outlier removal feature which removes the noisy data points is also included to achieve more accurate point cloud data. The evaluations on KITTY and ModelNet40 datasets demonstrate that proposed RANDT method achieves RMSE values of 0.054 m and 0.062 m with an error reduction of 18-25% even with noise and partial overlaps, and also attains highest point density uniformity score 0.92 than other baseline methods.