Smart Pothole Detection Using Deep Learning Based on Dilated Convolution

Area of Science:

• Computer Vision
• Deep Learning
• Transportation Engineering

Background:

• Potholes pose significant risks to transportation infrastructure and economic activity.
• Automated pothole detection using computer vision is crucial for efficient road maintenance.
• Existing methods require optimization for accuracy, speed, and cost-effectiveness.

Purpose of the Study:

• To develop efficient deep learning models for real-time pothole detection.
• To improve the accuracy and speed of automated pothole detection systems.
• To propose a cost-effective and easily implementable solution for road infrastructure monitoring.

Main Methods:

• Development of efficient deep learning Convolutional Neural Networks (CNNs).
• Proposal of a modified VGG16 (MVGG16) network with reduced layers and varied dilation rates.
• Integration of MVGG16 as a backbone for the Faster R-CNN architecture.
• Comparative performance analysis of YOLOv5 and Faster R-CNN with various backbones (ResNet101, ResNet50(FPN), VGG16, MobileNetV2, InceptionV3, MVGG16).

Main Results:

• The YOLOv5 Small (Y_s) model demonstrated superior speed for real-time pothole detection.
• Faster R-CNN with the proposed MVGG16 backbone achieved higher mean precision and reduced inference time compared to other backbones (VGG16, InceptionV3, MobileNetV2).
• The MVGG16 network effectively balances pothole detection accuracy and processing speed.

Conclusions:

• Deep learning models, particularly optimized CNNs, are highly effective for automated pothole detection.
• The modified VGG16 backbone enhances the performance of Faster R-CNN for pothole detection tasks.
• The proposed approach offers a promising solution for real-time, accurate, and efficient road infrastructure monitoring.