Interpolating and denoising point cloud data for computationally efficient environment modeling

Abstract

Light detection and ranging (LIDAR) is an important component of autonomous vehicles for environment modeling in real time. LIDAR generates a point cloud by analyzing the echo of light pulses scattered from the objects surrounding the train. Since the generated point clouds are too large to be used in practical applications, they need to be converted to a simpler and more compact form, while preserving all of the important features of the environment. Moreover, LIDAR signals are often affected by various noises or interferences, which rapidly decrease the signal-to-noise ratio of the signal with increasing distance. This study proposes a multi-scale analysis (MSA) strategy for interpolation and denoising of LIDAR signals efficiently. The developed method uses a coarse-to-fine hierarchical approximation that incrementally fits LIDAR signals up to a desired degree of accuracy. It is shown on three different signal types that the local analysis of the proposed method can eliminate noise more effectively than filtering and radial basis function based reconstruction of a signal provides computational efficiency compared to wavelet decomposition.