Scholarlink

<p>Detecting moving objects especially under drastic illumination changes and extreme weather conditions is highly challenging. To address this problem, we incorporate event cameras together with RGB cameras to achieve high dynamic range and low-latency perception, and we propose a Sparse SpatioTemporal moving-object detection Network (SSTNet). By performing multi-timescale event aggregation and adopting a sparsity-aware, adaptive feature-fusion strategy, the proposed network enables robust moving object detection under extreme conditions. Specifically, we design an Event Aggregation Module (EAM) based on ConvLSTM to effectively exploit the temporal information embedded in event streams. In addition, we introduce a Multi-level Aggregation Attention module (MAA) to achieve adaptive fusion of dual-modal features. MAA integrates features via a cross-attention mechanism and removes redundant information within the fused representations. Extensive experiments conducted on the DSEC-MOD dataset demonstrate that the proposed SSTNet achieves state-of-the-art performance in moving object detection.</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

Multi-level Aggregation Attention RGB–Event Fusion Network for Moving Object Detection

  • 期刊论文

  • 发表日期：

    2026年5月5日

查看全文
  • Yizhuo Fu,
  • Weiping Yang,
  • Zhiyong Zhang

摘要

Detecting moving objects especially under drastic illumination changes and extreme weather conditions is highly challenging. To address this problem, we incorporate event cameras together with RGB cameras to achieve high dynamic range and low-latency perception, and we propose a Sparse SpatioTemporal moving-object detection Network (SSTNet). By performing multi-timescale event aggregation and adopting a sparsity-aware, adaptive feature-fusion strategy, the proposed network enables robust moving object detection under extreme conditions. Specifically, we design an Event Aggregation Module (EAM) based on ConvLSTM to effectively exploit the temporal information embedded in event streams. In addition, we introduce a Multi-level Aggregation Attention module (MAA) to achieve adaptive fusion of dual-modal features. MAA integrates features via a cross-attention mechanism and removes redundant information within the fused representations. Extensive experiments conducted on the DSEC-MOD dataset demonstrate that the proposed SSTNet achieves state-of-the-art performance in moving object detection.