LiDAR Blooming Artifacts Estimation Method Induced by Retro-Reflectance with Synthetic Data Modeling and Deep Learning

This paper proposes a new method of blooming artifacts estimation, which is inherent in flash light detection and ranging (LiDAR) based on single-photon avalanche diode (SPAD), including the following two steps: (1) synthetic data generation by modeling the physical properties of a specific type of crosstalk noise (i.e., blooming effect); (2) a deep learning-based segmentation network to estimate per-pixel probability, indicating how likely a depth value is distorted by this noise. First, we captured around 17K images of depth and corresponding intensity data utilizing sequential flash type LiDAR sensor by driving around Seoul, South Korea. Then, we modeled the crosstalk noise due to retro-reflective signs by analyzing its unique physical properties. Finally, a deep learning-based segmentation network is trained and applied to estimate blooming artifact pixels. We demonstrate the feasibility of modeling blooming artifacts and the effectiveness of this synthetic data generation.