Real-Time Pothole Segmentation and Area Estimation Using Improved U-Net Models on Edge Devices
TL;DR
A pothole area measurement system based on U-Net models that employs the segmentation approach was created and deployed on edge computing devices and found that the improved ResNet U-Net model yielded the best results.
Abstract
In this study, a pothole area measurement system based on U-Net models that employs the segmentation approach was created and deployed on edge computing devices. The improved Dense U-Net, U-Net with dropout, ResNet U-Net, and standard U-Net models utilized hyperparameter optimization through random search. Shape feature extraction is used to convert the pothole area from pixels to area units. Various settings of pothole images were employed in the modeling procedure during the training and validation stages. It was found that the improved ResNet U-Net model yielded the best results. The ResNet U-Net model performance exhibited 0.6831 IoU, 0.8445 precision, 0.8222 recall, and 0.3233 loss. This model was installed on a Raspberry Pi edge computing device to segment potholes and quantify their area and diameter. The findings are considered acceptable and can be used in a road damage detection system implemented on edge devices.