Shadow-Based Vehicle Detection in Urban Traffic.

Area of Science:

• Computer Vision
• Automotive Safety Systems
• Artificial Intelligence

Background:

• Vehicle detection is crucial for Forward Collision Avoiding Systems (FACS).
• Existing vision-based methods often use a two-stage approach: hypotheses generation and verification.
• The initial stage of generating reliable vehicle hypotheses is critical for overall system performance.

Purpose of the Study:

• To develop a robust feature-based method for on-road vehicle detection in urban environments.
• To improve the hypotheses generation stage for vision-based vehicle detection.
• To enhance the performance and robustness of the initial stage of FACS.

Main Methods:

• A feature-based approach focusing on shadow detection under vehicles.
• Utilizing pixel properties and vertical intensity gradients caused by road shadows.
• Implementing a novel intensity thresholding strategy to define shadow upper bounds, reducing false positives.
• Employing morphological discrimination for candidate verification.

Main Results:

• The proposed method successfully generates hypotheses for vehicle candidates based on under-vehicle shadows.
• The refined thresholding strategy significantly reduces false positive rates compared to traditional methods.
• Promising detection performance and robustness were observed in daytime conditions, varying weather, and cluttered scenarios.

Conclusions:

• The developed method provides a reliable first stage for vehicle detection in FACS.
• Shadow analysis, with an improved thresholding technique, offers an effective way to generate vehicle hypotheses.
• The findings support the validation of this approach for enhancing automotive safety systems.