The Road Surface Anomalies Detection using Deep Convolutional Neural Networks with Transfer Learning Technique

Abstract

The road surface anomaly detection is an important task in road maintenance and road transportation to safety assurance for all road users. This research study the road surface anomalies detection using deep convolutional neural networks with transfer learning technique. The primary objective was to compare the performance of 5 deep convolutional neural networks architectures that are the pre-trained models, such as 1) Faster R-CNN_NAS, 2) Faster R-CNN Inception_ResNetV2 Atrous, 3) SSD ResNet50 FPN, 4 ) SSD MobileNetV1 FPN, and 5) Faster R-CNN ResNet101. The images dataset of road surface anomalies in the natural environment for training and testing have 6 categories, such as 1) pothole 2) longitudinal cracks, 3) transverse cracks, 4) alligator cracks, 5) patch, and 6) speed bumps.  The results show that the SSD ResNet50 FPN model provides the highest mean average precision of 87.38% and least test times. The study provides practical methodology and models that could be applied to automatically detect road surface anomalies in other regions to support road informatics, improve quality, efficiency and driving safety.