Modelling and estimating occluded regions : from 3D LiDAR point clouds

Sammanfattning

Occlusions and limited sensor coverage are unavoidable challenges for autonomous drivingsystems operating in real-world environments. Regions that are not directly observable by onboard sensors may contain obstacles or vulnerable road users, making safe decision-making difficult. This thesis investigates how occluded and unseen regions can be modelled andestimated using probabilistic spatial representations. The work reviews common environment representations used in robotic perception, highlighting the limitations of point clouds and low-dimensional maps in reasoning about occlusions and incomplete information. Based on this analysis, a probabilistic three-dimensional occupancy mapping framework is selected, where unseen regions are represented through uncertainty and occlusions correspond to persistent unknown areas. A baseline occupancy mapping approach and a custom probabilistic voxel-based method are implemented and evaluated using LiDAR data from real-world driving scenarios. The proposed method adopts an octree-inspired volumetric representation at a fixed resolution, maintaining occupancy probabilities at the voxel level to represent free, occupied and unknown space. The analysis shows how occupancy estimates evolve as LiDAR data is accumulated, where repeated measurements reinforce stable structures while regions with limited sensor coverage remain unresolved. Regions with inconsistent observations retain intermediate occupancy values, indicating temporal variability in the environment. The probabilistic formulation captures both spatial structures and sensor visibility constraints, resulting in a map that reflects how information is gathered and refined across successive observations. Overall, the proposed approach provides a consistent and interpretable description of environments with occlusions and limited sensor range. By maintaining uncertainty in unobserved regions, the method supports cautious reasoning and forms a suitable basis for autonomous driving tasks such as navigation and risk-aware decision-making.