Spatial scene temporal behavior framework for anomaly detection

doi:10.1016/j.dsp.2025.105076

Spatial scene temporal behavior framework for anomaly detection

2025-05-01

0

PRE

AI

李洪均 (Hongjun Li) *

章国安 (Guoan Zhang)

DOI：10.1016/j.dsp.2025.105076

Original

Original request for help

Save

Abstract

En 中文

Unsupervised video anomaly detection plays an important role in public security. However, it is a challenging task in real scene due to the following reasons: i) incomplete data on normal patterns and behaviors; and ii) the weak adaptability to new scenes. In this paper, we present a spatial scene temporal behavior (SSTB) framework, which takes into account both the few-shot learning and scene transference to anomaly detection. To this end, we first introduce a transformation network of scene feature generation (Tnet) to learn detailed scene information. Moreover, we present content loss and style loss at the feature level to maximize the consistency of scene generation. Additionally, a siamese network of temporal behavior detection (Snet) is designed for different objects, where coarse features and fine-grained features are sequentially extracted to represent temporal behaviors on normal and pseudo-abnormal sequences. As a result, our method achieves anomaly detection by combining the results from scene feature generation and temporal behavior detection. Experimental results indicate that our method outperforms the existing approaches on three benchmark datasets.

Keywords:

Anomaly detection

Unsupervised framework

Temporal behavior

Scene features

Spatial scene temporal behavior framework for anomaly detection

Abstract

Journal

Organization