An Automated Detection Method for Motor Vehicles Encroaching on Non-Motorized Lanes Based on Unmanned Aerial Vehicle...

Area of Science:

• Computer Vision
• Traffic Engineering
• Artificial Intelligence

Background:

• Motor vehicle encroachment into non-motorized lanes is a frequent urban violation.
• Existing detection methods using overhead views suffer from perspective distortion, occlusion, and jitter, leading to unreliable evidence.
• Unmanned aerial vehicles (UAVs) offer a promising monitoring perspective but require robust detection systems.

Purpose of the Study:

• To develop a cascaded system for accurate detection and evidence generation of motor vehicle encroachment violations from UAV perspectives.
• To overcome limitations of traditional methods sensitive to perspective distortion and occlusion.
• To provide a reliable system for traffic governance and enforcement.

Main Methods:

• A two-stage detection cascade using lightweight vehicle detection and a YOLOv5n-based oriented bounding box (OBB) detector for license plates.
• U-Net lane region segmentation for pixel-level spatial constraints.
• PP-OCRv4 with perspective rectification for enhanced license plate recognition.
• Temporal stability using an N-frame counter and area ratio criterion.

Main Results:

• Improved detection accuracy from 67.0% to 92.0% on a controlled benchmark.
• Reduced false positive rate from 32.35% to 5.88% compared to a baseline.
• The system effectively localizes vehicles, segments lanes, recognizes tilted license plates, and stabilizes alarms.

Conclusions:

• The proposed cascaded UAV-view system significantly enhances the accuracy and reliability of detecting motor vehicle encroachment.
• The system provides both violation alarms and verifiable license plate evidence, suitable for practical traffic governance.
• This approach addresses key challenges in aerial traffic monitoring, offering a robust solution for enforcement.