Page 76 - 2025S

UEC Int’l Mini-Conference No.54 69

Anomaly Classification with Scene Description in Public Roads

Jose Emilio VERA CORDERO ∗1 and TAKAHASHI Hiroki 2

1 UEC Exchange Study Program (JUSST Program)
2 Department of Informatics
The University of Electro-Communications, Tokyo, Japan

Keywords: Abnormal, Anomaly, Anomaly Detection, Anomaly Classification, Surveillance System.

Abstract
Anomalous events are those that deviate from common or expected behavior. Detecting such events
in public spaces, such as streets and avenues, is essential for automating surveillance systems and ensur-
ing safety. However, a simple alarm triggered by anomaly detection is often insufficient, as it typically
lacks any explanation or justification regarding the detected event. Therefore, accurate classification
of anomalous events has become an increasingly important task. Proper classification not only enables
a better understanding of the anomalies detected but also facilitates more effective and context-aware
responses. Several studies have shown that incorporating textual information can provide deeper se-
mantic understanding, which significantly improves the accuracy of anomaly classification. In this work,
we propose a system for anomaly classification that combines visual features (from video) and textual
features (from event descriptions), leveraging the CLIP model for this purpose. We employ videos
from the widely known large-scale UCF-Crime dataset, along with the UCA dataset, which provides
fine-grained descriptions of all events occurring in the UCF-Crime videos.

∗
The author is supported by JASSO Scholarship.