Scholarlink

<p>As autonomous driving technology advances, the need for precise environmental perception grows. While onboard cameras are commonly used for real-time monitoring, their limited placement and viewing angles leave blind spots. This is particularly problematic in complex traffic environments, where hidden vehicles may be missed, increasing the risk of traffic accidents. Even with multi-view sharing through roadside cameras, large differences in perspectives and difficulty in aligning object features often result in insufficient detection accuracy for blind spot vehicles. To address these issues, this paper proposes the Sparse Attention and Graph Matching Network (SAGM-Net), which consists of two modules: the Sparse Attention Module (SAM) and the Graph Matching Module (GMM). SAM significantly reduces the computational cost on high-resolution images through a sparse self-attention mechanism in local windows, and enhances the ability to capture local spatial structures by incorporating relative positional encoding. This effectively addresses challenges related to diverse vehicle appearances and occlusions. GMM iteratively optimizes the matching probabilities between cross-view nodes within a multi-scale graph structure. By utilizing a message-passing strategy, it achieves precise alignment of features between onboard and roadside views, thereby enhancing the capture and recognition of blind spot vehicle features. Experimental results show that SAGM-Net achieves outstanding performance on multi-angle, multi-scenario vehicle-road cooperative datasets. Compared with existing methods, it significantly improves the detection accuracy of blind spot vehicles and demonstrates higher robustness and adaptability in complex road environments. The code is available at github.com/yanqzhang/Sparse-Attention-and-Graph-Matching-Network.</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

Sparse attention and graph matching network for multi-view vehicle detection in collaborative perception systems

  • 期刊论文

  • 发表日期：

    2025年7月30日

查看全文
  • Yanqin Zhang,
  • Zhongliang Wang

摘要

As autonomous driving technology advances, the need for precise environmental perception grows. While onboard cameras are commonly used for real-time monitoring, their limited placement and viewing angles leave blind spots. This is particularly problematic in complex traffic environments, where hidden vehicles may be missed, increasing the risk of traffic accidents. Even with multi-view sharing through roadside cameras, large differences in perspectives and difficulty in aligning object features often result in insufficient detection accuracy for blind spot vehicles. To address these issues, this paper proposes the Sparse Attention and Graph Matching Network (SAGM-Net), which consists of two modules: the Sparse Attention Module (SAM) and the Graph Matching Module (GMM). SAM significantly reduces the computational cost on high-resolution images through a sparse self-attention mechanism in local windows, and enhances the ability to capture local spatial structures by incorporating relative positional encoding. This effectively addresses challenges related to diverse vehicle appearances and occlusions. GMM iteratively optimizes the matching probabilities between cross-view nodes within a multi-scale graph structure. By utilizing a message-passing strategy, it achieves precise alignment of features between onboard and roadside views, thereby enhancing the capture and recognition of blind spot vehicle features. Experimental results show that SAGM-Net achieves outstanding performance on multi-angle, multi-scenario vehicle-road cooperative datasets. Compared with existing methods, it significantly improves the detection accuracy of blind spot vehicles and demonstrates higher robustness and adaptability in complex road environments. The code is available at github.com/yanqzhang/Sparse-Attention-and-Graph-Matching-Network.