AMCW time-of-flight scanning LiDAR based on parallel-phase demodulation and its image enhancements

Abstract

In this dissertation, a novel 3D amplitude-modulated continuous wave (AMCW) coaxial scanning LiDAR based on parallel-phase demodulation is proposed and demonstrated. Specifically, to enable highly precise depth measurement with high optical signal-to-noise ratio (SNR) and short integration time, the parallel-phase demodulation method is combined with the single-pixel coaxial scanning optics in this proposed LiDAR. According to the experimental validation, the proposed 3D scanning LiDAR shows highly precise full-HD (FHD) resolution 3D depth image with illumination power lower than 30 mW and 800 nsec integration time. Meanwhile, to enhance the 3D depth image quality, suppression of depth error caused by multipath interference (MPI) is also conducted in this dissertation. To generate precise synthetic training dataset representing the MPI phenomena of the proposed AMCW coaxial scanning LiDAR, a physical simulation model which combines light transport equations and various sensor noise equations is designed. Using this simulation, multiple MPI synthetic data for various scenarios are synthesized at multiple frequencies to be trained using Bayesian optimized-extreme gradient boost (XGBoost) ensemble. Subsequently, the trained XGBoost ensemble is used to post-correct the measured depth images with MPI. According to the experimental validation results, the mean absolute error (MAE) of distance error due to MPI could be reduced to lower than 3 mm in indoor measurement condition. In addition to MPI suppression, the internal stray light generated inside the coaxial optics is also suppressed based on Gaussian-mixture model (GMM) and particle swarm optimization (PSO). Specifically, since the ratio of directly reflected light from object to the internal stray light varies with the reflectivity of measured object, the measured depth value at the same object distance can differ depending on the object’s reflectivity. To correct such depth error, images of calibration checkboard are measured at multiple distances and used to estimate the parameters of internal stray light which make the post-corrected checkboard depth maps flat. Such estimated internal stray light parameters are also used to correct the other complex image scenes distorted by the same internal stray light. According to the experimental validation using the calibration checkboard, the average L1 loss due to stray light could be reduced to 3.2 mm based on the proposed method. At last, this dissertation proposes a 20 tap-based pulse waveform parallel-phase demodulation method to minimize the depth error caused by strong external light. The relevant validation results are also presented. For the future works, the proposed LiDAR and depth error correction methods will be utilized with object segmentation/classification algorithm for the 3D recognition of autonomous robots, bin picking robot, and smart security systems.