Arrayed LiDAR signal analysis for automotive applications
MetadataShow full item record
Light detection and ranging (LiDAR) is one of the enabling technologies for advanced driver assistance and autonomy. Advances in solid-state photon detector arrays offer the potential of high-performance LiDAR systems but require novel signal processing approaches to fully exploit the dramatic increase in data volume an arrayed detector can provide. This thesis presents two approaches applicable to arrayed solid-state LiDAR. First, a novel block independent sparse depth reconstruction framework is developed, which utilises a random and very sparse illumination scheme to reduce illumination density while improving sampling times, which further remain constant for any array size. Compressive sensing (CS) principles are used to reconstruct depth information from small measurement subsets. The smaller problem size of blocks reduces the reconstruction complexity, improves compressive depth reconstruction performance and enables fast concurrent processing. A feasibility study of a system proposal for this approach demonstrates that the required logic could be practically implemented within detector size constraints. Second, a novel deep learning architecture called LiDARNet is presented to localise surface returns from LiDAR waveforms with high throughput. This single data driven processing approach can unify a wide range of scenarios, making use of a training-by-simulation methodology. This augments real datasets with challenging simulated conditions such as multiple returns and high noise variance, while enabling rapid prototyping of fast data driven processing approaches for arrayed LiDAR systems. Both approaches are fast and practical processing methodologies for arrayed LiDAR systems. These retrieve depth information with excellent depth resolution for wide operating ranges, and are demonstrated on real and simulated data. LiDARNet is a rapid approach to determine surface locations from LiDAR waveforms for efficient point cloud generation, while block sparse depth reconstruction is an efficient method to facilitate high-resolution depth maps at high frame rates with reduced power and memory requirements.