Abstract:

In view of the problem of cluster target detection in low-altitude maritime environments， this paper proposed a multi-feature fusion detection network based on gated cross-attention and bidirectional feature pyramid （GCAB-MFF Detector） to address the challenges brought by complex sea conditions， non-Gaussian clutter interference， and the nonlinear characteristics of radar echoes. The network extracted features in the target range dimension and the range-Doppler dimension respectively through a dual-channel architecture and utilized the gated cross-attention module （GCAF） to model cross-modal feature correlations； furthermore， it introduced the bidirectional feature pyramid （BiFPN） to achieve efficient integration of multi-level features through weighted multi-scale fusion and semantic consistency optimization and ultimately output the target existence detection and cluster scale classification results simultaneously. Experimental results indicate that on the self-built dataset， the detection rate of the GCAB-MFF detector reaches 99.92%， and the scale classification accuracy is 92.56%. These results are significantly improved compared to the traditional CFAR method （detection rate < 89%） and traditional classification networks （classification accuracy < 86%）， which verifies its robustness and practicality in low-altitude maritime scenarios.

Key words: cluster target detection, deep learning, feature fusion, low-altitude maritime target, dual-channel network, attention mechanism

摘要：

针对海上低空环境下集群目标检测问题，提出一种门控交叉注意力与双向特征金字塔的多特征融合检测网络（GCAB-MFF detector），以应对复杂海况、非高斯杂波干扰及雷达回波非线性特性带来的挑战。该网络通过双通道架构分别提取目标距离维与距离-多普勒维特征，并利用门控交叉注意力模块（GCAF）建模跨模态特征关联；进一步引入双向特征金字塔（BiFPN），通过加权多尺度融合与语义一致性优化，实现多层次特征的高效整合，最终同步输出目标存在性检测与集群规模分类结果。实验结果表明，在自建数据集上，GCAB-MFF检测器检测率达99.92%，规模分类准确率为92.56%，较传统CFAR方法（检测率<89%）、传统分类网络（分类精度<86%）有显著提升，验证了其在海上低空场景下的鲁棒性与实用性。

关键词: 集群目标检测, 深度学习, 特征融合, 海上低空目标, 双通道网络, 注意力机制