A Detection Transformer-Based Intelligent Identification Method for Multiple Types of Road Traffic Safety Facilities.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Manual detection of traffic safety facilities (TSFs) is unsustainable due to their complexity and variety.
• Automating TSF detection is crucial for efficient traffic management and maintenance.

Purpose of the Study:

• To develop an automated system for detecting diverse road traffic safety facilities.
• To establish a comprehensive dataset for training and evaluating TSF detection models.

Main Methods:

• Construction of the TSF-CQU dataset with 8,410 instances across six TSF types from car-recorder imagery.
• Implementation and comparison of detection models: Denoising Anchor Box (DINO), Faster R-CNN, and Yolov7.
• Evaluation using performance metrics including mean average precision (mAP) and average precision (AP).

Main Results:

• The DINO model achieved the highest performance on the TSF-CQU dataset with an mAP of 82.2%.
• Individual AP values for most TSFs exceeded 0.8, with streetlights at 0.77 and rods at 0.648.
• The DINO model demonstrated high accuracy, with minimal erroneous recognition, validating the contrastive denoising training approach.

Conclusions:

• The DINO model is highly effective for automatic detection of road traffic safety facilities.
• The TSF-CQU dataset provides a valuable resource for advancing research in this domain.
• Further refinement may address the few instances of missed detections observed with the DINO model.