DeepCPD：WiFi CSI信号的儿童检测低成本方案

DeepCPD：WiFi CSI信号的儿童检测方案

论文信息

• 标题： DeepCPD: Deep Learning Based In-Car Child Presence Detection Using WiFi
• 来源： IEEE Journal of Selected Areas in Sensors, 2025
• 链接： https://arxiv.org/abs/2505.01973

---

WiFi感知技术原理

CSI信号基础

# WiFi CSI感知原理
CSI_SENSING_PRINCIPLE = {
    "csi_definition": "Channel State Information（信道状态信息）",
    "measurement": "WiFi信号在子载波上的幅度和相位",
    "dimensions": ["时间", "子载波", "发射天线", "接收天线"],
    
    "sensing_mechanism": {
        "static_environment": "CSI稳定",
        "human_presence": "人体反射导致CSI变化",
        "vital_signs": "呼吸/心跳引起周期性CSI变化"
    },
    
    "advantages": [
        "利用现有WiFi硬件（无需额外传感器）",
        "穿透遮挡物（毯子/座椅）",
        "隐私保护（无图像）",
        "全光照条件工作"
    ],
    
    "challenges": [
        "分辨率低（vs mmWave雷达）",
        "环境噪声干扰",
        "WiFi标准差异"
    ]
}

CPD检测流程

graph LR
    A[WiFi设备] -->|发送信号| B[CSI采集]
    B -->|预处理| C[去噪+归一化]
    C -->|特征提取| D[深度学习模型]
    D -->|分类| E{儿童检测}
    E -->|检测到| F[触发警告]
    E -->|无检测| G[继续监测]

---

核心算法详解

CSI预处理

import numpy as np
from typing import Tuple

class CSIPreprocessor:
    """
    CSI信号预处理器
    
    去噪、归一化、特征提取
    """
    
    def __init__(self,
                 n_subcarriers: int = 56,
                 n_tx: int = 2,
                 n_rx: int = 2):
        self.n_subcarriers = n_subcarriers
        self.n_tx = n_tx
        self.n_rx = n_rx
        
    def preprocess(self,
                   csi_raw: np.ndarray,
                   window_sec: float = 5.0,
                   sample_rate: float = 100.0) -> np.ndarray:
        """
        预处理CSI数据
        
        Args:
            csi_raw: 原始CSI数据 (T, N_sub, N_tx, N_rx, 2)
            window_sec: 分析窗口（秒）
            sample_rate: 采样率
            
        Returns:
            csi_features: 预处理后的特征图
        """
        window_samples = int(window_sec * sample_rate)
        
        # 提取幅度
        amplitude = np.abs(csi_raw[-window_samples:, :, :, :, 0] + 
                         1j * csi_raw[-window_samples:, :, :, :, 1])
        
        # 去除直流分量
        amplitude = amplitude - np.mean(amplitude, axis=0, keepdims=True)
        
        # 归一化
        amplitude = amplitude / (np.std(amplitude) + 1e-6)
        
        # 滤波（低通滤波去除高频噪声）
        from scipy.signal import butter, filtfilt
        b, a = butter(4, 0.3, btype='low')
        amplitude_filtered = filtfilt(b, a, amplitude, axis=0)
        
        return amplitude_filtered
    
    def extract_features(self, csi_clean: np.ndarray) -> np.ndarray:
        """
        提取特征
        
        Args:
            csi_clean: 清洗后的CSI数据
            
        Returns:
            features: 特征向量
        """
        features = []
        
        # 时域特征
        features.append(np.mean(csi_clean, axis=0).flatten())
        features.append(np.std(csi_clean, axis=0).flatten())
        features.append(np.max(csi_clean, axis=0).flatten())
        features.append(np.min(csi_clean, axis=0).flatten())
        
        # 频域特征（FFT）
        fft_result = np.fft.fft(csi_clean, axis=0)
        features.append(np.abs(fft_result).mean(axis=0).flatten())
        
        # 组合特征
        features = np.concatenate(features)
        
        return features


class DeepCPD_Model(nn.Module):
    """
    DeepCPD深度学习模型
    
    基于CSI特征检测儿童存在
    """
    
    def __init__(self,
                 input_dim: int = 1024,
                 hidden_dim: int = 256):
        super().__init__()
        
        self.classifier = nn.Sequential(
            nn.Linear(input_dim, hidden_dim),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(hidden_dim, 128),
            nn.ReLU(),
            nn.Dropout(0.3),
            nn.Linear(128, 64),
            nn.ReLU(),
            nn.Linear(64, 2)  # [无儿童, 有儿童]
        )
        
    def forward(self, features: torch.Tensor) -> torch.Tensor:
        """
        前向传播
        
        Args:
            features: CSI特征向量 (B, input_dim)
            
        Returns:
            logits: 分类输出 (B, 2)
        """
        return self.classifier(features)

---

性能对比

方案	成本	准确率	遮挡穿透	隐私
mmWave雷达	$25-35	94%	✅	✅
WiFi CSI	$5-10	88%	✅	✅
摄像头	$15-20	96%	❌	❌
超声波	$10-15	75%	⚠️	✅

---

IMS开发落地指导

# WiFi CPD部署方案
WIFI_CPD_DEPLOYMENT = {
    "hardware": {
        "wifi_chip": "Intel 5300 / Atheros AR9300",
        "antenna": "MIMO 2×2",
        "cost": "$5-10（利用现有WiFi）"
    },
    
    "software": {
        "csi_extraction": "Linux CSI Tool",
        "model_size": "500KB",
        "inference_time": "30ms"
    },
    
    "suitable_vehicles": [
        "经济型车型",
        "已有WiFi模块车型",
        "成本敏感方案"
    ]
}

---

总结： WiFi CSI是CPD的低成本替代方案，适合成本敏感车型。虽然准确率略低于mmWave雷达（88% vs 94%），但成本优势明显（$5-10 vs $25-35），且隐私友好。