A Study on Depth Estimation and Digital Terrain Model Reconstruction for Mobile Mapping Systems

Mobile Mapping Systems (MMSs) are advanced platforms for collecting precise geospatial data. They use technologies like LiDAR and digital imaging to gather detailed environmental information. MMSs have two main components: georeferencing and mapping. Georeferencing aligns digital data with coordinate systems using Global Navigation Satellite Systems (GNSS) for high accuracy. In areas where GNSS is unavailable, Simultaneous Localization and Mapping (SLAM) technology is used to map unknown environments and track MMS locations. The accuracy and density of depth maps are crucial for SLAM performance, affecting the system's ability to create maps and track positions.

This dissertation addresses depth estimation challenges in MMSs by introducing the Double-stage Adaptive Refinement Scheme (DARS). DARS is designed to improve depth estimation in dynamic environments and can be integrated with pre-trained networks. It can also be extended to Panoramic DARS (PanoDARS) for SLAM applications using panoramic images, improving the HDPV-SLAM system's performance by addressing LiDAR depth sparsity and depth association issues. The thesis also explores reconstructing Digital Terrain Models (DTMs) from Digital Surface Models (DSMs), essential for accurate mapping within MMSs. DTMs provide elevation data for creating geospatial products like orthophotos, topographic maps, and 3D urban models. To achieve precise DTM reconstruction, a geospatially induced autoencoder called SB-SUBNET is proposed. This autoencoder uses two geospatial inductive biases that leverage the relationship DTM = DSM - nDSM. The first bias reconstructs the DTM by subtracting the network's output (nDSM) from its input (DSM). The second bias is a subtractive skip connection that integrates this geospatial information directly into the network, enhancing performance using inherent geospatial relationships in the data.

While SB-SUBNET shows promising results, it alters all DSM values, including terrain values, affecting DTM accuracy. To address this, a new multi-task learning approach named DB-SUBNET was developed. This approach segments non-ground points and performs regression specifically for these points while preserving ground points. This method improves DTM reconstruction accuracy and ensures the resulting DTMs provide a reliable foundation for high-precision geospatial products. Enhanced DTM reconstruction is vital for creating detailed and accurate maps, supporting urban planning, infrastructure development, environmental monitoring, and disaster management.