Assessing UAV-derived Point Cloud Accuracy in Street-Level Mapping Networks

Thesis Abstract

This study addresses the persistent challenge of achieving reliable geometric integrity in street-level mapping produced from Unmanned Aerial Vehicle (UAV) derived point clouds, where urban clutter, occlusions, and sensor fusion inconsistencies often degrade positional accuracy and feature delineation. The aim is to quantify and improve the spatial accuracy of UAV-derived point clouds within street-scale networks, with a focus on how flight parameters, sensor configuration, and data processing workflows influence accuracy metrics. Specific objectives are to (i) quantify vertical and horizontal errors of UAV point clouds relative to high-precision reference data, (ii) evaluate the impact of flight altitude, overlap, and ground control distribution on point cloud quality, (iii) assess the influence of LiDAR-enabled and photogrammetric multi-sensor fusion workflows on street-level feature representation, and (iv) develop a recommended processing pipeline that optimizes accuracy for street networks under constrained urban environments. The study adopts an empirical field design conducted in a mid-sized urban corridor spanning 4 km of street frontages with varied building geometries. A population of fixed urban blocks is identified, and a stratified sampling approach yields 20 representative street segments where data collection occurs. The UAV platform comprises a high-resolution RGB camera and a lightweight LiDAR scanner mounted on a quadcopter, coupled with ground control points measured by a Differential GNSS system to ensure centimetre-level georeferencing. Data collection involves capture campaigns at three altitude strata (60 m, 100 m, 140 m) with forward and lateral overlaps of 75% and 65% respectively, and a standardized ground control network of 40 evenly distributed points per segment. Reference datasets consist of Terrestrial Laser Scanning (TLS) scans acquired with a high-precision Phase-Shifted TLS instrument and surveyed control targets, serving as the ground-truth benchmark. Instruments include the UAV sensor suite for point cloud generation, TLS-based reference datasets for accuracy assessment, a DGPS/GNSS receiver for controlling georeferencing, and a weather-monitoring station to capture atmospheric variables that may affect sensor performance. Data processing follows a two-pronged workflow (i) photogrammetric processing using structure-from-motion and multi-view stereo (MVS) to generate dense point clouds, and (ii) LiDAR point cloud processing with calibrated sensor fusion to produce integrated multi-sensor datasets. The validity and reliability of processing outputs are ensured through standardized calibration procedures, repeatability tests across campaigns, and cross-validation against TLS data. Analytical techniques include descriptive statistics to summarize error distributions, root-mean-square error (RMSE) and mean absolute error (MAE) calculations for horizontal and vertical components, and regression analyses to identify relationships between accuracy metrics and experimental factors (altitude, overlap, ground control density, sensor fusion settings). ANOVA tests determine significant differences across altitude strata and processing workflows. A mixed-methods component incorporates qualitative assessment of feature fidelity (e.g., curb lines, façade edges) via expert scoring to complement quantitative accuracy metrics. Expected findings indicate that lower flight altitudes with dense ground control networks yield the highest horizontal and vertical accuracy, with RMSEs anticipated around 3–5 cm for horizontal and 4–6 cm for vertical when TLS is used as reference, while higher altitudes degrade accuracy by 15–25% depending on occlusions. Multi-sensor fusion workflows are expected to improve vertical integrity for complex street canyons by reducing systematic biases observed in photogrammetric-only approaches. The study contributes to knowledge by delivering a calibrated, replicable processing pipeline tailored to street-level mapping that explicitly accounts for urban geometry and sensor fusion trade-offs, and by providing empirically validated guidelines for flight planning in similar urban environments. The theoretical framing engages ontological and epistemological considerations of sensor fusion accuracy within geospatial information theory, referencing error propagation models and the bias-variance trade-off as explanatory mechanisms for observed differences between workflows. The main conclusion asserts that UAV-derived point clouds can achieve centimetre-level street-scale accuracy under carefully designed flight campaigns and robust sensor fusion, provided that ground control density and urban occlusion mitigation are prioritized. Recommendations include adopting a standardized ground control network scheme, implementing adaptive flight planning to maintain consistent detail in obstructed street sections, and integrating LiDAR-augmented photogrammetry as a preferred workflow in dense urban corridors. Implications for practice emphasize improved asset inventory, safer road-network modeling, and enhanced interoperability between UAV-derived datasets and existing GIS infrastructures.