Scholarlink

<p>Detecting rare events, such as fights or falls, in surveillance videos is critical for public safety, but is complicated by data imbalance, occlusions, and complex spatiotemporal dynamics. We present a Hybrid TokenShift-Stochastic Transformer that combines a 3D convolutional backbone with a transformer attention mechanism to address these issues. The token shift module effectively captures the temporal context by shifting feature channels across frames, whereas a stochastic Local Winner-Takes-All (LWTA) layer enhances the sparsity for robust feature selection. The model achieved area under the curve-receiver operating characteristic (AUC-ROC) values of 96.8%, 90.5%, and 95.2% on the Hockey Fight, TNUE-Fight, and UCF-Crime datasets, respectively. Additionally, it outperforms the baseline models in terms of computational cost, demonstrating a 26% reduction in FLOPs while maintaining real-time processing at 15.2ms and utilizing 2.3GB of memory. These advancements make this method highly suitable for real-time surveillance in dynamic environments.</p>

错误：搜索内容不能为空，请输入英文关键词
错误：关键词超出字数限制，请精简
高级检索

Hybrid TokenShift-stochastic transformer for rare event detection in video surveillance

  • 期刊论文

  • 发表日期：

    2025年11月28日

查看全文
  • Yahaya Idris Abubakar,
  • Mamadou Dia,
  • Patrick Siarry,
  • Alice Othmani

摘要

Detecting rare events, such as fights or falls, in surveillance videos is critical for public safety, but is complicated by data imbalance, occlusions, and complex spatiotemporal dynamics. We present a Hybrid TokenShift-Stochastic Transformer that combines a 3D convolutional backbone with a transformer attention mechanism to address these issues. The token shift module effectively captures the temporal context by shifting feature channels across frames, whereas a stochastic Local Winner-Takes-All (LWTA) layer enhances the sparsity for robust feature selection. The model achieved area under the curve-receiver operating characteristic (AUC-ROC) values of 96.8%, 90.5%, and 95.2% on the Hockey Fight, TNUE-Fight, and UCF-Crime datasets, respectively. Additionally, it outperforms the baseline models in terms of computational cost, demonstrating a 26% reduction in FLOPs while maintaining real-time processing at 15.2ms and utilizing 2.3GB of memory. These advancements make this method highly suitable for real-time surveillance in dynamic environments.