Development of a method to detect anomaly of civil infrastructure after the earthquake with existing CCTV cameras

Abstract

Following an earthquake, the government’s disaster management bureau immediately seeks to clarify the degree of damage to infrastructure. Closed Circuit Television (CCTV) cameras are widely deployed to monitor infrastructure on roads and rivers. However, when detecting anomalies by comparing pre- and post-earthquake images, the lack of post-earthquake training images poses a the fundamental challenge. Furthermore, varying lighting conditions and moving objects (e.g., vehicles) often induce false positives. In this paper, to address these unsupervised learning challenges, we propose a novel method utilizing Optimal Transport (OT) in addition to PatchCore, a state-of-the-art anomaly detection algorithm. OT theory enables us to compare the feature distributions between pre- and post-earthquake images without incurring excessive computational overhead.

To evaluate our method, we conducted experiments in three stages. First, we extracted the feature distributions from the 2nd and 3rd layers of the PatchCore backbone with approximate 1000 normal images at noon for each of 8 CCTV cameras under various surrounding environments. Second, across 10 simulated post-earthquake timeframes, we calculated the cosine similarity using the sinkhorn algorithm of OT. Third, we visualized the anomaly score on a heat map with and without OT. In this way, we validated the false positive of OT. Moreover, we validated false negatives with 3 cases from 2 additional CCTV cameras, which captured the actual damage to infrastructure. As a result, we demonstrated that the method with OT can reduce the anomaly degree of false areas.