Road Damage Detection Acquisition System based on Deep Neural Networks for Physical Asset Management

TL;DR

This paper introduces a large-scale, balanced road damage dataset and evaluates traditional and deep learning object detection methods for effective damage classification and assessment in infrastructure management.

Contribution

It provides a new comprehensive dataset of road damages and compares multiple detection models, including deep learning approaches, for practical deployment.

Findings

Deep learning models outperform traditional methods in accuracy.

MobileNet and RetinaNet achieve a good balance of speed and performance.

The dataset enables standardized benchmarking for road damage detection.

Abstract

Research on damage detection of road surfaces has been an active area of re-search, but most studies have focused so far on the detection of the presence of damages. However, in real-world scenarios, road managers need to clearly understand the type of damage and its extent in order to take effective action in advance or to allocate the necessary resources. Moreover, currently there are few uniform and openly available road damage datasets, leading to a lack of a common benchmark for road damage detection. Such dataset could be used in a great variety of applications; herein, it is intended to serve as the acquisition component of a physical asset management tool which can aid governments agencies for planning purposes, or by infrastructure mainte-nance companies. In this paper, we make two contributions to address these issues. First, we present a large-scale road damage dataset, which includes a more balanced and representative set of damages. This dataset is composed of 18,034 road damage images captured with a smartphone, with 45,435 in-stances road surface damages. Second, we trained different types of object detection methods, both traditional (an LBP-cascaded classifier) and deep learning-based, specifically, MobileNet and RetinaNet, which are amenable for embedded and mobile and implementations with an acceptable perfor-mance for many applications. We compare the accuracy and inference time of all these models with others in the state of the art.