Identification and Tracking of Vehicles between Multiple Cameras on Bridges Using a YOLOv4 and OSNet-Based Method

Area of Science:

• Engineering
• Computer Science
• Transportation

Background:

• Vehicle load estimation is crucial for bridge structural health monitoring (SHM).
• Existing methods like bridge weight-in-motion (BWIM) lack vehicle location data.
• Multi-camera vehicle tracking on bridges is challenging due to non-overlapping fields of view.

Purpose of the Study:

• To develop a robust computer vision system for vehicle detection and tracking across multiple, non-overlapping camera views on bridges.
• To improve the accuracy of vehicle location and movement data for comprehensive bridge SHM.
• To enable the acquisition of temporal-spatial vehicle load distribution on entire bridges.

Main Methods:

• A You Only Look Once v4 (YOLOv4) and Omni-Scale Net (OSNet)-based model was employed for vehicle detection and tracking.
• A modified Intersection over Union (IoU)-based tracker incorporated vehicle appearance and bounding box overlap for intra-camera tracking.
• The Hungary algorithm facilitated vehicle matching across different camera video feeds.

Main Results:

• The proposed method achieved 97.7% accuracy for single-camera vehicle tracking.
• Multi-camera vehicle tracking accuracy exceeded 92.5%.
• A dedicated dataset of 25,080 images featuring 1727 vehicles was created for model training and evaluation.

Conclusions:

• The developed system effectively tracks vehicles across multiple cameras, overcoming visual field limitations.
• This technology significantly enhances the ability to monitor vehicle loads and their spatial-temporal distribution on bridges.
• The findings contribute to more accurate and comprehensive bridge structural health monitoring.