Geo-Context Aware Study of Vision-Based Autonomous Driving Models and Spatial Video Data.

Area of Science:

• Computer Vision
• Geospatial Analysis
• Machine Learning for Autonomous Systems

Background:

• Deep learning (DL) models are advancing autonomous driving by learning from vast video datasets.
• Current methods predict driving behaviors from on-vehicle camera data but lack integrated geospatial context.

Purpose of the Study:

• To develop a geo-context aware visualization system for analyzing autonomous driving model (ADM) predictions.
• To integrate DL model performance with geospatial visualization for a comprehensive study of driving behaviors.

Main Methods:

• Developed a visualization system combining DL model performance metrics with geospatial visualization techniques.
• Integrated city-wide and street-level analysis of multiple DL models' prediction behaviors.
• Incorporated road images and video content for detailed visual exploration.

Main Results:

• The system allows users to study DL model performance alongside geospatial attributes on map views.
• Users can discover and compare prediction behaviors of different DL models at various geographical scales.
• Demonstrated the utility and effectiveness of the visualization system through use cases and expert evaluation.

Conclusions:

• The developed system offers a novel visual exploration platform for autonomous driving DL model designers.
• Enhanced understanding of ADM predictions by integrating visual data with geographical context.
• Facilitates improved development and validation of autonomous driving technologies.