Vehicle Detection and Attribution from a Multi-Sensor Dataset Using a Rule-Based Approach Combined with Data Fusion

Area of Science:

• Remote Sensing
• Computer Vision
• Geospatial Analysis

Background:

• Vehicle detection is crucial for national security, disaster relief, and traffic monitoring.
• Existing methods often use single data sources, limiting comprehensive analysis.
• Multi-data fusion offers potential for more robust vehicle detection.

Purpose of the Study:

• To develop and evaluate a novel, primarily LiDAR-based method for stationary vehicle detection.
• To enhance vehicle detection accuracy by integrating RGB/MSI imagery and road network data.
• To assign detailed attributes to detected vehicles, including 3D, relational, and spectral properties.

Main Methods:

• A LiDAR-centric approach for stationary vehicle detection.
• Fusion of LiDAR point clouds with RGB/MSI imagery and vector road network data.
• Utilized datasets from Houston, TX (IEEE GRSS 2018) and Vaihingen, Germany (ISPRS), plus DIRSIG-simulated data.

Main Results:

• Achieved 92% precision and 92% recall on the Houston dataset with 1476 ground truth vehicles.
• Demonstrated improved classification and reduced false positives through data fusion.
• Successfully assigned various attributes and height profiles to detected vehicles.

Conclusions:

• The developed data fusion method effectively detects stationary vehicles with high accuracy.
• Integrating multiple data sources significantly enhances the reliability of vehicle detection systems.
• Limitations include false positives from low vegetation and challenges with closely spaced or low-density point cloud vehicles.