Video Anomaly Detection Through Structured Track and Pose Representation and Hierarchical Transformer

The general contributions of this paper are mentioned below.
• We develop a structured anomaly detection system that
utilizes human pose features, object information and a
track-aware fixed-slot scene representation system.
• We propose an empty-score-based object allocation sys
tem that can preserve slot-to-ID consistency even when
objects are temporarily missing from consecutive frames.
• We introduce a multiscale hierarchical transformer that
first captures short-term temporal patterns within 10
frame chunks and then learns long-term patterns using the
resulting chunk-level representations, requiring 31.96%
fewer MACs than the flat baseline.