Related Experiment Video
Updated: Jan 18, 2026
03:31
Author Spotlight: Enhancement of Salient Object Detection for Smart Grid Applications
Published on: December 15, 2023
1.0K
Dual-Stream Attention-Enhanced Memory Networks for Video Anomaly Detection
Weishan Gao1, Xiaoyin Wang1, Ye Wang1
1China Aerospace Academy of Systems Science and Engineering, Beijing 100048, China.
Sensors (Basel, Switzerland)
|September 13, 2025
Summary
This study introduces a new method for weakly supervised video anomaly detection (WSVAD) to improve accuracy. The novel approach enhances feature representation and discrimination, significantly reducing false alarms and improving detection of unusual events.
Area of Science:
- Computer Vision
- Artificial Intelligence
- Machine Learning
Background:
- Weakly supervised video anomaly detection (WSVAD) struggles with complex temporal dependencies and distinguishing similar events.
- Existing methods often yield high false alarm rates due to background noise interference.
Purpose of the Study:
- To develop a novel WSVAD method that enhances feature representation and discrimination for more accurate anomaly detection.
- To address challenges in modeling temporal dynamics and mitigating background noise.
Main Methods:
- Employs a hierarchical multi-scale temporal encoder and position-aware global relation network for robust temporal representations.
- Utilizes a dual-stream attention-enhanced memory network with bidirectional spatial attention for precise event discrimination and noise reduction.
Main Results:
- Achieved state-of-the-art performance on UCF-Crime (87.43% AUC) and XD-Violence (85.51% AP) datasets using only RGB features.
- Demonstrated significant improvements in detecting salient events and reducing false positives.
Conclusions:
- The proposed "attention-guided prototype matching" paradigm effectively tackles key WSVAD challenges.
- The method enables robust and precise anomaly detection in videos.