Three-Dimensional Outdoor Pedestrian Road Network Map Construction Based on Crowdsourced Trajectory Data

Jianbo Tang, Tianyu Zhang, Junjie Ding, Ke Tao, Chen Yang, Jianbing Xiang, Xia NingISPRS Int. J. Geo Inf 2025 · DOI

TL;DR

This paper introduces a novel hierarchical method for constructing detailed 3D pedestrian road networks from crowdsourced GPS trajectory data. By addressing challenges like low elevation accuracy and complex outdoor environments, it aims to provide richer navigation data than existing 2D or less detailed 3D approaches.

What Problem Does This Paper Address?

Creating accurate and up-to-date outdoor pedestrian road network maps is surprisingly difficult. Traditional methods like field surveys are expensive and time-consuming, while existing automated approaches often fall short. Vehicle trajectory-based methods, while mature for 2D road extraction, struggle to capture the crucial 3D information (like elevation and slope) vital for pedestrian navigation. Furthermore, outdoor environments present unique challenges such as tree occlusion and complex terrain, which can degrade GPS accuracy, especially in the vertical dimension. This gap limits the development of sophisticated pedestrian navigation applications.

Methods

The paper proposes a hierarchical "decomposition-combination" strategy for 3D pedestrian road network construction. The core steps are:

  1. Trajectory Data Preprocessing: This crucial first step addresses the inherent inaccuracies in raw GPS data. It involves cleaning, smoothing, and importantly, trajectory data aggregation using a gravity-repulsion force model to improve spatial accuracy. A key innovation here is trajectory elevation information rectification based on Digital Elevation Model (DEM) data, which significantly enhances vertical accuracy.
  2. Identification of Intersections and Pedestrian Areas: This stage leverages the distinct movement patterns of pedestrians. Trajectory points are analyzed for their directionality. An improved DBSCAN clustering algorithm that considers direction constraints is employed to identify grid cells corresponding to road segments, intersections, and pedestrian areas. This method accounts for both spatial proximity and directional similarity.
  3. Hierarchical Extraction of the 3D Pedestrian Road Network: This is the main contribution.
  4. Intersection Scope and Entrance/Exit Point Identification: For identified intersection grids, the method refines their boundaries and pinpoints entrance/exit points using trajectory clustering and density analysis.
  5. 3D Road Centerline Extraction: Trajectories within and between identified areas are clustered. A three-dimensional B-spline fitting method is then used to generate smooth 3D road centerlines from these clusters. This approach is designed to handle complex intersections and adjacent roads that might be merged by simpler methods.
  6. Topological Extraction and Optimization: The extracted road centerlines, intersections, and pedestrian areas are fused to create a complete 3D pedestrian road network graph.

Datasets

The study utilizes crowdsourced pedestrian trajectory data from two sources:

  • 2bulu: 1985 trajectory data points.
  • foooooot: 1421 trajectory data points.

These datasets were collected from the Yuelu Mountain Scenic Area in Changsha, Hunan Province, China, between January 2015 and June 2023. The data includes user ID, longitude, latitude, elevation, timestamp, and speed, totaling approximately 6.2 million 3D trajectory points. The study also uses Advanced Land Observing Satellite (ALOS) Digital Elevation Model (DEM) data with a spatial resolution of 12.5 m for elevation correction.

Key Insights and Ideas

  • Elevation Accuracy is Key for Pedestrian Navigation: Unlike vehicle navigation, pedestrian routes are heavily influenced by elevation changes (slopes, stairs). The paper's focus on improving elevation accuracy using DEM data is a critical insight.
  • Movement Patterns Reveal Road Structure: Pedestrian trajectory directionality is a strong indicator of road type (straight road, intersection, open area). Leveraging this directional information in clustering is more robust than purely spatial methods.
  • Hierarchical Approach for Complexity: Outdoor environments are complex. A hierarchical method that first identifies broad areas (intersections, pedestrian zones) and then refines the road structure within them is more effective than a monolithic approach.
  • 3D B-Spline Fitting for Smoothness and Accuracy: Using B-spline fitting in 3D allows for the generation of smooth road centerlines that accurately represent the trajectory data, even for curved or complex paths.

Key Results

  • The proposed method achieved higher accuracy in extracting 3D road information compared to existing methods.
  • Quantitative evaluation for intersection identification showed a precision of 92.2%, recall of 67.8%, and an F1 score of 0.781. This outperforms the MAMC algorithm (F1=0.705) and the KDE algorithm (F1=0.574) in terms of precision and F1 score, especially in areas with uneven trajectory distribution.
  • The method effectively identified pedestrian areas, with results overlaid on remote sensing images showing good correspondence.
  • The average plane position deviation of extracted road centerlines was approximately 4.4 m, meeting the requirements for outdoor navigation applications.
  • The 3D road network construction method demonstrated an ability to accurately extract multiple roads that are adjacent in the plane but separated vertically, a capability lacking in step-by-step methods like the improved KDE algorithm.

Limitations and Future Work

  • Data Sparsity: In areas with missing trajectory data or rarely visited trails, the method may fail to obtain complete road information or accurately detect intersections due to insufficient directional clustering.
  • Future Work: The authors suggest integrating multi-source spatio-temporal data (e.g., remote sensing images, social media data) to overcome data sparsity issues and enable finer extraction and dynamic updates of the road network's geometric, topological, and semantic information.

Summary updated

BibTeX
@article{tang2025three,
  author = {Jianbo Tang and Tianyu Zhang and Junjie Ding and Ke Tao and Chen Yang and Jianbing Xiang and Xia Ning},
  title = {{Three-Dimensional Outdoor Pedestrian Road Network Map Construction Based on Crowdsourced Trajectory Data}},
  journal = {ISPRS Int. J. Geo Inf.},
  year = {2025},
  doi = {10.3390/ijgi14040175},
}
Figure