Efficient Visual Analytics of Localized High-Frequency Multimodal All-Traffic Trajectory Data at Urban Intersections

Zihan Liu, Jiaying Xue, Yu Feng, Tumasch Reichenbacher, Liqiu MengAbstracts 2025 · DOI

Introduction

Understanding the complex dynamics of urban intersections is critical for improving traffic safety and efficiency. Traditional methods of capturing traffic data often fall short of providing a comprehensive view of all road users. This paper introduces a visual analytics framework designed to address the "3H" challenges—High density, High interactivity, and High dynamics—inherent in off-board trajectory data. By utilizing roadside LiDAR, the proposed system moves beyond tracking individual GPS-equipped vehicles to capturing the high-frequency movements and interactions of all road users, including pedestrians and cyclists.

The Shift to Off-Board Sensing

Traditional trajectory data collection typically relies on on-board sensors (e.g., GPS) or mixed sensors (e.g., Bluetooth/Wi-Fi). While these methods provide valuable insights, they are limited to capturing only the subset of objects equipped with sensors. To achieve a holistic understanding of urban environments, off-board sensors such as roadside LiDAR and cameras are required. These sensors can track every object within their range at high frequencies, though they introduce significant data management challenges.

Comparison of sensor categories for trajectory data collection, highlighting the advantages of off-board sensing for full object coverage.

The authors identify three core "3H" challenges associated with this data:

  1. High Spatio-Temporal Density: Full coverage at high sampling rates creates massive datasets.
  2. High Interactivity: Objects react to one another and the environment (e.g., traffic signals).
  3. High Dynamics: Rapid changes in movement patterns, such as sudden braking, lead to unpredictable data keyframes.

Methodology

The proposed visual analytics system employs six strategies to transform raw data into actionable insights:

  • Semantic Modeling: Maps raw coordinates to the intersection's "streetscape" (lanes, sidewalks, crosswalks) to provide environmental context.
  • Semantic & Attribute Filtering: Enables users to filter data by category (e.g., cyclists) or complex semantic labels.
  • Keyframe Identification: Automatically detects critical behavioral moments, such as the initiation of deceleration.
  • Pattern Clustering: Groups objects based on keyframes to discover movement trends within specific categories.
  • Multi-Dimensional Interaction Analysis: Analyzes kinematics (speed/distance), behavior (yielding), and safety (risk levels).
  • Perspective Switching: Allows toggling between a global overview and focused views of specific modal interactions.

Dataset and Case Study

The system was evaluated using a real-world dataset collected from a roadside LiDAR station in the Maxvorstadt district of Munich, Germany. The dataset comprises over five hours of peak-time traffic recorded at a frequency of 10 Hz. Trajectories are stored as 3D bounding boxes, each defined by nine parameters: 3D centroid coordinates, box dimensions, and rotation angles.

Key Insights and Results

A central finding of this research is that raw trajectory points gain meaning only when contextualized within a semantic map. For instance, identifying a coordinate as being on a sidewalk versus a vehicle lane is crucial for near-miss analysis. Furthermore, by focusing on "critical timestamps" or keyframes rather than the entire 10Hz stream, the system significantly reduces cognitive load for analysts while highlighting the most important behavioral events.

The system demonstrates the ability to:

  • Identify Bottlenecks: Pinpoint areas where traffic flow is impeded.
  • Improve Safety: Highlight "near-crash" events often missed by traditional accident reports.
  • Evaluate Interventions: Provide data-driven assessments of new infrastructure or signal timings.
  • High-Fidelity Tracking: Capture nuanced behaviors, such as pedestrian hesitation, through high-frequency LiDAR data.

Conclusion and Future Work

While the current system is a prototype focused on a single intersection, it lays the groundwork for advanced urban design research. Future work will address the computational challenges of scaling the system to city-wide networks and further automating keyframe identification for increasingly complex behavioral events.

Summary updated

BibTeX
@inproceedings{liu2025efficient,
  author = {Zihan Liu and Jiaying Xue and Yu Feng and Tumasch Reichenbacher and Liqiu Meng},
  title = {{Efficient Visual Analytics of Localized High-Frequency Multimodal All-Traffic Trajectory Data at Urban Intersections}},
  booktitle = {Abstracts},
  year = {2025},
  doi = {10.5194/ica-abs-10-178-2025},
}
Figure