Scholarlink

<p>In modern video surveillance systems, identifying unusual events and retrieving specific moments from extensive video footage are critical yet challenging tasks. These challenges stem from the complexity of dynamic scenes and the need for accurate, query-based video understanding. To address this, we propose the query- attentive spatiotemporal transformer for moment recognition (QuAST-MR) a novel framework built on the tthese limitations, we transformer architecture that enhances the retrieval of relevant video segments in response to user-defined queries. QuAST-MR integrates temporal self-attention masking with query-guided feature learning, enabling precise temporal localization. The model uses a ResNet-50 backbone for visual feature extraction and has been evaluated on two publicly available benchmark datasets UAV-VID and VIRAT which simulate real-world surveillance conditions. Our framework achieves a 2.5% improvement in mean Reciprocal Rank (MRR) on UAV-VID and a 3.12% gain in Normalized Discounted Cumulative Gain (NDCG) on VIRAT. These metrics assess the relevance and ranking quality of retrieved video segments, reflecting the model’s effectiveness in returning correct results earlier and more accurately. These improvements demonstrate the practical potential of QuAST-MR for moment retrieval in realistic surveillance contexts, contributing toward more intelligent and responsive video analysis systems.</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

Enhancing Anomaly Detection and Moment Retrieval with a Query-Driven Transformer Framework

  • 期刊论文

  • 发表日期：

    2026年6月24日

查看全文
  • Pratibha Singh,
  • Alok Kumar Singh Kushwaha

摘要

In modern video surveillance systems, identifying unusual events and retrieving specific moments from extensive video footage are critical yet challenging tasks. These challenges stem from the complexity of dynamic scenes and the need for accurate, query-based video understanding. To address this, we propose the query- attentive spatiotemporal transformer for moment recognition (QuAST-MR) a novel framework built on the tthese limitations, we transformer architecture that enhances the retrieval of relevant video segments in response to user-defined queries. QuAST-MR integrates temporal self-attention masking with query-guided feature learning, enabling precise temporal localization. The model uses a ResNet-50 backbone for visual feature extraction and has been evaluated on two publicly available benchmark datasets UAV-VID and VIRAT which simulate real-world surveillance conditions. Our framework achieves a 2.5% improvement in mean Reciprocal Rank (MRR) on UAV-VID and a 3.12% gain in Normalized Discounted Cumulative Gain (NDCG) on VIRAT. These metrics assess the relevance and ranking quality of retrieved video segments, reflecting the model’s effectiveness in returning correct results earlier and more accurately. These improvements demonstrate the practical potential of QuAST-MR for moment retrieval in realistic surveillance contexts, contributing toward more intelligent and responsive video analysis systems.