Automated Ice Road Surface Detection: Anomaly-Based Unique Dataset Construction Tailored to Individual CCTV Cameras

Abstract

The objective of this study is to propose and evaluate a method for automatically constructing a training dataset specific to each closed-circuit television (CCTV) camera to identify ice-road surface images using anomaly detection. In this approach, dry-road surface images were defined as normal, while ice-road surface images were considered as anomalous. By training the anomaly detection model solely on normal images, the system can identify ice-road surface images. However, the accuracy of identification may decline owing to variations in shooting locations, individual CCTV camera installation sites, and environmental conditions. To address these challenges, we used data from the Automated Meteorological Data Acquisition System (AMeDAS) near CCTV cameras to automatically select normal images, thereby enabling the construction of camera-specific training datasets. Specifically, if precipitation and temperature data from the nearest AMeDAS station—often located several kilometers away—consistently remained below predefined thresholds for a certain duration, the corresponding CCTV images were assumed to depict dry-road surfaces. We applied the proposed method to CCTV camera images and experimentally evaluated it using the PatchCore anomaly detection technique. The results demonstrated high accuracy, achieving an area under the receiver operating characteristic score of 0.961 and 0.999 during the daytime and nighttime, respectively, validating its effectiveness in detecting ice-road surface images. To demonstrate the practical viability of the proposed method, we validated it using footage from two additional CCTV cameras. The results showed that ice-road surface images could be easily identified using the proposed approach.