收藏切换
RDT-MoE: Integrating Redundancy Reduction and Mixture-of-Experts for Violence Action Recognition
收藏切换
Jianfeng Yang1, Dawei Yu2, Rui Cao1, *, Jiahang Ren1, Ping Wang1, Qiyao Hu3
Journal of Computer Research and Development | 2026, 63(8) : 2127 - 2142
Less
收藏切换
Journal of Computer Research and Development | 2026, 63(8): 2127-2142
• Artificial Intelligence •
RDT-MoE: Integrating Redundancy Reduction and Mixture-of-Experts for Violence Action Recognition
Full
Jianfeng Yang1, Dawei Yu2, Rui Cao1, *, Jiahang Ren1, Ping Wang1, Qiyao Hu3
Affiliations
  • 1College of Computer Science, Northwest University, Xi’an 710127
  • 2Shanghai Huawei Technologies Co., Shanghai 201206
  • 3School of Electronic Information, Northwest University, Xi’an 710127
  • born in 1981. Master, senior engineer, master supervisor. His main research interests include software development methodologies, and data science and its applications in higher education

    born in 1999. Master, engineer. His main research interest includes action recognition

    born in 1991. PhD, lecturer, master supervisor. Member of CCF. His research interests include text representation learning, application optimization of generative large language models, and diffusion-based model artistic generation

    born in 2002. Master candidate. His main research interests include large model development and application, and network security

    born in 2002. Master candidate. His main research interest includes large model visual reasoning

    born in 1992. PhD, engineer, master supervisor. Member of IEEE. Her main research interests include AI art and multimodal generation

About Author:

born in 1981. Master, senior engineer, master supervisor. His main research interests include software development methodologies, and data science and its applications in higher education

born in 1999. Master, engineer. His main research interest includes action recognition

born in 1991. PhD, lecturer, master supervisor. Member of CCF. His research interests include text representation learning, application optimization of generative large language models, and diffusion-based model artistic generation

born in 2002. Master candidate. His main research interests include large model development and application, and network security

born in 2002. Master candidate. His main research interest includes large model visual reasoning

born in 1992. PhD, engineer, master supervisor. Member of IEEE. Her main research interests include AI art and multimodal generation

Published: 2026-08-01 doi: 10.7544/issn1000-1239.202550743
Outline
收藏切换

To tackle the dual challenges of spatiotemporal feature redundancy and limited generalization in violent action recognition, we present RDT-MoE, a lightweight yet effective framework that couples redundancy-diminishing token (RDT) with a mixture-of-experts (MoE) architecture. The RDT module employs spatially and temporally decoupled attention to disentangle motion from appearance, and uses a dynamic Top-k token selection strategy to suppress nearly 50% of background redundancy, keeping only the most salient motion cues for subsequent encoding. Building on these compact tokens, we develop a LoRA-guided expert system and train it with a two-stage strategy, where we first jointly optimize the encoder and LoRA adapters to learn strong shared representations and stable adaptation signals, and then freeze the backbone to focus on refining the gating function and expert composition, which further improves robustness, convergence stability, and generalization across diverse scenes. To address limitations of existing benchmarks, we also introduce diverse violence action dataset (DVAD), a higher-resolution, multi-view violence dataset that is 50% larger than RWF-2000 and better reflects real-world variability. Extensive evaluations on RWF-2000 and DVAD, together with component-wise ablations, show that RDT-MoE outperforms VideoMAE by 2.00% and 2.34% in accuracy, respectively, while reducing parameters by about 50%, demonstrating strong efficiency, generalization, and practical applicability. This design supports deployment on resource-limited surveillance devices and maintains stable performance across diverse scenes consistently.

violence action recognition  /  violence action dataset  /  LoRA  /  MoE  /  pre-training
杨建锋, 于大为, 曹瑞, 任佳航, 王平, 胡琦瑶. RDT-MoE:融合冗余削减与混合专家的暴力行为识别方法. 计算机研究与发展, 2026 , 63 (8) : 2127 -2142 . DOI: 10.7544/issn1000-1239.202550743
Jianfeng Yang, Dawei Yu, Rui Cao, Jiahang Ren, Ping Wang, Qiyao Hu. RDT-MoE: Integrating Redundancy Reduction and Mixture-of-Experts for Violence Action Recognition[J]. Journal of Computer Research and Development, 2026 , 63 (8) : 2127 -2142 . DOI: 10.7544/issn1000-1239.202550743
Year 2026 volume 63 Issue 8
87
42
Cite this Article
BibTeX
Article Info
doi: 10.7544/issn1000-1239.202550743
  • Receive Date:2025-10-14
  • Online Date:2026-11-17
  • Published:2026-08-01
Article Data
Affiliations
History
  • Received:2025-10-14
  • Revised:2026-03-25
  • Accepted:2026-04-03
Affiliations
    1College of Computer Science, Northwest University, Xi’an 710127
    2Shanghai Huawei Technologies Co., Shanghai 201206
    3School of Electronic Information, Northwest University, Xi’an 710127
References
Share
https://castjournals.cast.org.cn/joweb/jsjyjyfz/EN/10.7544/issn1000-1239.202550743
Share to
QR

Scan QR to access full text

Cite this article
BibTeX
Citations
表12种不同金属材料的力学参数
科
Family
属数
Number of
genus
种数
Number of
species
占总种数比例
Percentage of
total species (%)
属
Genus
种数
Number of
species
占总种数比例
Percentage of total
species (%)
鹅膏菌科Amanitaceae 2 11 5.26 鹅膏菌属 Amanita 10 4.78
小菇科 Mycenaceae 2 12 5.74 丝盖伞属 Inocybe 5 2.39
多孔菌科 Polyporaceae 8 14 6.70 蜡蘑属 Laccaria 5 2.39
红菇科 Russulaceae 3 23 11.00 小皮伞属 Marasmius 6 2.87
小菇属 Mycena 11 5.26
光柄菇属 Pluteus 5 2.39
红菇属 Russula 17 8.13
栓菌属 Trametes 5 2.39
关闭全屏
  • BibTeX
  • EndNote
  • RefWorks
  • TxT