Development of a Large-Scale Roadside Facility Detection Model Based on the Mapillary Dataset.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Transportation Engineering

Background:

• Accurate detection of road facilities is crucial for high-definition (HD) maps and intelligent transportation systems (ITSs).
• Deep learning object detection offers improved accuracy and efficiency for HD map reconstruction and advanced driver-assistance systems (ADASs).
• Existing research often focuses on single-class object detection, necessitating a broader comparison for multi-class road facility detection.

Purpose of the Study:

• To systematically compare the performance of three recent deep learning algorithms: Mask R-CNN, YOLOx, and YOLOv7.
• To evaluate these algorithms for large-scale, multi-class road facility detection using the Mapillary dataset.
• To assess the generalization ability of trained models on a custom Japanese road environment dataset.

Main Methods:

• Implementation and comparison of Mask R-CNN, YOLOx, and YOLOv7 algorithms.
• Evaluation metrics included recall, precision, mean F1-score, and computational consumption.
• Testing model performance on the Mapillary dataset and a custom Japanese road dataset.

Main Results:

• YOLOv7 demonstrated superior performance in road facility detection, achieving 87.57% precision and 72.60% recall.
• Models trained on the Mapillary dataset showed significant generalization capabilities when tested on the Japanese road environment.
• The study provides a comprehensive comparison of the strengths and limitations of the evaluated networks.

Conclusions:

• YOLOv7 is highly effective for large-scale, multi-class road facility detection.
• The findings support the use of deep learning for enhancing HD maps and ITSs.
• The generalization ability of models trained on diverse datasets is confirmed, paving the way for robust autonomous driving systems.