Bayesian image reconstruction and adaptive scene sampling in single-photon LiDAR imaging

Abstract

Three-Dimensional multispectral Light Detection And Ranging (LiDAR) used with time-correlated Single-Photon (SP) detection has emerged as a key imaging modality for high-resolution depth imaging due to its high sensitivity and excellent surface-to-surface resolution. This allowed depth imaging through adversarial conditions with a prime role in numerous applications. However, several practical challenges currently limit the use of LiDAR in real-world conditions. Large data volume constitutes a major challenge for multispectral SP-LiDAR imaging due to the acquisition of millions of events per second that are usually gathered in large histogram cubes. This challenge is more evident when the useful signal photons are attenuated and the background noise is amplified as a result of imaging through a scattering environment such as underwater or fog. Another limitation includes the detection of multiple-surfaces-per pixel which usually occurs when imaging through semi-transparent materials (e.g., windows, camouflage), or in long-range profiling. This thesis proposes robust and fast computational solutions to improve the acquisition and processing of LiDAR data while measuring uncertainty on high-dimensional data. A smart task-based sampling framework is proposed to improve the acquisition process and reduce data volume. In addition, the processing was improved using a Bayesian approach to different types of inverse problems (e.g. spectral classification, and scene reconstruction). The contributions of this thesis enables fast and robust 3D reconstruction of complex scenes, paving the way for the extensive use of single-photon imaging in real-world applications.