Field study toward anomaly road damage detection with drive recorder

Abstract

As one of the ways to reduce road maintenance costs, road damage detection with a mobile camera is gaining attention. Most of conventional damage detection use supervised learning, nevertheless three practical drawbacks exist. Firstly, supervised learning requires a high manual cost to collect annotated data for training. Secondly, some damages are rarely observed, resulting in imbalanced data and difficulty in training an efficient model for all damage categories. Additionally, annotators may not identify such rare damages correctly. Thirdly, supervised learning cannot detect unknown categories of damages, though unknown categories are often found in a practical scene. To overcome these three drawbacks, we propose an ensemble model that combines anomaly detection and supervised damage detection. Anomaly detection can detect previously unknown and rare types of damage, while supervised damage detection ensures damages frequently observed on roads. Two different models cover wider categories of road damages. Our ensemble model is expected to achieve higher accuracy and lower manual cost.