Scalable computational methods for 3D reconstruction using single-photon Lidar : towards online approaches

Abstract

Lidar systems allow the reconstruction of 3D scenes at high resolution. They have recently received considerable interest across a wide domain of applications. The technology consists of illuminating a target with laser pulses and measuring the time of arrival of the pulses that have been reflected on it to estimate both its distance and reflectivity. However, the analysis of Lidar measurements can be considerably challenged by additional detection events that arise from external light sources such as ambient illumination. Algorithms have been developed to ensure efficient 3D reconstructions in the presence of such spurious detection events. The growing development of technological applications based on the Lidar technology requires efficient computational methods. The goal of this thesis is to develop advanced algorithms for single-photon Lidar that can address the problems mentioned above, namely how to ensure robustness and an appropriate computational complexity for adaptive processing. First, a link is established with sampling theory by highlighting that the reconstruction problem can be viewed as restoring a signal having a finite rate of innovation. Since the presence of spurious detection events in Lidar measurements complicates the approach, a denoising algorithm is proposed to reduce the effect of external light sources before addressing reconstruction problems. The 3D reconstruction problem is then addressed from a recent type of multispectral Lidar where the photon returns associated with different wavelengths are concatenated into a single histogram. The proposed Bayesian approaches differ from that considered with the sampling method as they enable the estimation of background illumination levels as well as the modelling of the observations with Poisson processes. Two methods are proposed to ensure satisfying estimation performance in a considerably reduced computational time compared to state-of-the-art methods. Finally, an online algorithm is proposed to address the reconstruction of a dynamic 3D scene. A pseudo-Bayesian framework is adopted where the classical likelihood term is replaced by a maximum divergence estimator. The proposed pseudo-Bayesian online algorithm enables close to real-time estimation of the model parameters and performs efficient 3D video reconstruction.