Mamba-Transformer Hybrid Network for Efficient Video Anomaly Detection
Year: 2026
Type: conference-paper
Abstract: To address the issues of insufficient long-range temporal dependency modeling, separated spatial-temporal feature learning, and high computational complexity in video anomaly detection, this paper proposes a Mamba-Transformer hybrid network. This model integrates the linear-complexity long-sequence modeling advantage of Mamba and the global semantic capture capability of Transformer, and achieves efficient spatial-temporal feature learning through multi-scale spatial-temporal feature extraction, dual-branch complementary modeling, and adaptive feature fusion. Meanwhile, lightweight optimization strategies are adopted to reduce computational cost, balancing detection accuracy and inference efficiency. Extensive experiments verify that the proposed method achieves excellent performance on mainstream datasets and can effectively meet the engineering deployment requirements of real-time video anomaly detection.
Institution: NARI Group (China)
Language: English
Subfield: Artificial Intelligence
Field: Computer Science
Domain: Physical Sciences
Open Access status: hybrid
OpenAlex ID: W7202139201