Data-Augmented Deep Learning Models for Abnormal Road Manhole Cover Detection.

Area of Science:

• Computer Vision
• Artificial Intelligence
• Road Safety Engineering

Background:

• Anomalous road manhole covers present significant road safety risks in urban environments.
• Deep learning models are crucial for automated detection of these anomalies in smart city initiatives.
• Current challenges include the scarcity of anomalous manhole cover data, hindering effective model training.

Purpose of the Study:

• To develop an advanced data augmentation method for improving road anomaly manhole cover detection models.
• To address the data scarcity issue in training deep learning models for road safety.
• To enhance the generalization capabilities of anomaly detection systems.

Main Methods:

• A novel data augmentation technique is proposed, utilizing synthetic samples not present in the original dataset.
• The method automatically determines optimal pasting positions for manhole cover samples.
• It predicts transformation parameters using visual prior experience and perspective transformations for realistic integration.

Main Results:

• The proposed method significantly improves the performance of anomaly detection models.
• A notable increase in mean average precision (mAP) of at least 6.8 was achieved compared to baseline models.
• The technique effectively captures the actual shape and context of manhole covers on roads.

Conclusions:

• The developed data augmentation approach offers a more accurate and efficient way to expand training datasets.
• This method enhances the robustness and accuracy of deep learning models for detecting anomalous manhole covers.
• The findings contribute to improving road safety through advanced computer vision applications in smart cities.