分布式雷达人体活动识别：ConvNeXt-MHSA-BiGRU 框架，87.56% 连续活动识别精度

论文：Robust Continuous Human Activity Recognition Using Deep Learning and Distributed Radar Sensors
arXiv：2609.10419
精度：87.56%（连续活动识别）

论文信息

项目 内容
arXiv 2609.10419
架构 ConvNeXt-MHSA-BiGRU
精度 87.56%
传感器 5 个分布式雷达
提升 +2.45% vs CNN-RNN

核心创新

  1. ConvNeXt + MHSA + BiGRU 三模块 — 空间特征提取 + 多头自注意力 + 时序建模
  2. 5 个分布式雷达传感器 — 多视角覆盖，解决单传感器遮挡问题
  3. 连续活动识别 — 不需要活动分割，直接连续输出
  4. 87.56% 平均精度 — 优于 CNN-RNN 基线 2.45%

方法

1
2
3
4
5
6
7
8
9
10
11
12
13
14
15
16
17
18
19
20
21
22
23
24
25
26
27
28
29
30
31
32
33
34
35
36
37
38
39
40
41
42
43
44
45
46
47
48
49
50
51
52
53
54
55
56
57
58
59
60
61
62
63
64
65
66
67
68
69
70
71
72
73
74
75
76
77
78
79
80
81
82
83
84
85
86
87
88
89
90
91
92
93
94
95
96
97
98
99
100
101
102
103
104
105
106
107
108
109
110
111
import torch
import torch.nn as nn

class DistributedRadarHAR(nn.Module):
    """分布式雷达人体活动识别"""
    
    def __init__(self, num_sensors=5, num_classes=10, 
                 hidden_dim=256, num_heads=8):
        super().__init__()
        
        # 每个雷达的 ConvNeXt 特征提取
        self.sensor_encoders = nn.ModuleList([
            ConvNeXtEncoder(out_dim=hidden_dim) 
            for _ in range(num_sensors)
        ])
        
        # 多传感器融合
        self.sensor_fusion = nn.MultiheadAttention(
            hidden_dim, num_heads, batch_first=True
        )
        
        # MHSA 时空注意力
        self.mhsa = nn.MultiheadAttention(
            hidden_dim, num_heads, batch_first=True
        )
        
        # BiGRU 时序建模
        self.bigru = nn.GRU(
            hidden_dim, hidden_dim // 2,
            num_layers=2, batch_first=True,
            bidirectional=True
        )
        
        # 分类头
        self.classifier = nn.Sequential(
            nn.Linear(hidden_dim, 128),
            nn.GELU(),
            nn.Dropout(0.2),
            nn.Linear(128, num_classes)
        )
        
    def forward(self, radar_sequence):
        """
        Args:
            radar_sequence: (B, T, S, 1, H, W)
                T=time, S=sensors, H×W=range-doppler map
        """
        B, T, S, _, H, W = radar_sequence.shape
        
        # 每个传感器编码
        sensor_feats = []
        for s in range(S):
            feat = self.sensor_encoders[s](
                radar_sequence[:, :, s]
            )  # (B, T, hidden_dim)
            sensor_feats.append(feat)
        
        # 多传感器注意力融合
        fused = torch.stack(sensor_feats, dim=2)  # (B, T, S, hidden_dim)
        fused_flat = fused.reshape(B * T, S, -1)
        
        attn_out, _ = self.sensor_fusion(
            fused_flat, fused_flat, fused_flat
        )
        fused = attn_out.mean(dim=1).reshape(B, T, -1)
        
        # MHSA 时空
        normed = nn.LayerNorm(fused.size(-1)).to(fused.device)(fused)
        mhsa_out, _ = self.mhsa(normed, normed, normed)
        fused = fused + mhsa_out
        
        # BiGRU 时序
        gru_out, _ = self.bigru(fused)
        
        # 取最后时间步
        out = gru_out[:, -1, :]
        
        return self.classifier(out)


class ConvNeXtEncoder(nn.Module):
    """简化 ConvNeXt 编码器"""
    
    def __init__(self, out_dim=256):
        super().__init__()
        self.encoder = nn.Sequential(
            nn.Conv2d(1, 32, 7, stride=2, padding=3),
            nn.BatchNorm2d(32), nn.GELU(),
            nn.Conv2d(32, 64, 3, stride=2, padding=1),
            nn.BatchNorm2d(64), nn.GELU(),
            nn.Conv2d(64, 128, 3, stride=2, padding=1),
            nn.BatchNorm2d(128), nn.GELU(),
            nn.AdaptiveAvgPool2d(1),
            nn.Flatten(),
            nn.Linear(128, out_dim)
        )
        
    def forward(self, x):
        B, T = x.shape[:2]
        x = x.reshape(B * T, *x.shape[2:])
        feat = self.encoder(x)
        return feat.reshape(B, T, -1)


if __name__ == "__main__":
    model = DistributedRadarHAR(num_sensors=5, num_classes=10)
    x = torch.randn(2, 30, 5, 1, 64, 64)  # B=2, T=30, S=5, 64x64 RD map
    out = model(x)
    print(f"Input: {x.shape}")
    print(f"Output: {out.shape}")
    print(f"Params: {sum(p.numel() for p in model.parameters()):,}")

IMS 应用启示

舱内多雷达活动识别

活动类型 雷达检测能力 Euro NCAP 相关
正常驾驶 ✅ 手臂微动 —
手机使用 ✅ 手持到耳边 D-02
转身看后座 ✅ 躯干旋转 —
前倾/异常姿态 ✅ 姿态变化 OOP
儿童微动 ✅ 呼吸/小动作 CPD

分布式雷达部署

1
2
[前挡风雷达] ─── [驾驶员侧雷达] ─── [副驾侧雷达]
                    [后座左雷达] ─── [后座右雷达]
雷达位置 型号 覆盖区域 用途
后视镜 IWR6843AOP 前排全域 DMS
A柱左 IWR1443 驾驶员侧 侧向姿态
A柱右 IWR1443 副驾侧 OOP
后视镜上方 IWR6843 后排 CPD
行李箱 IWR6843 后排全域 CPD

参考文献

  1. Hiam, S.M. & Soleymani, M. “Robust Continuous Human Activity Recognition Using Deep Learning and Distributed Radar Sensors.” arXiv:2609.10419, 2026.

本文由 AI 研究助手自主生成，基于 arXiv 2609.10419 论文解读。