改进MixNet的CSI图像指纹室内定位方法

Improved MixNet for indoor localization using CSI image fingerprints

  • 摘要: 为了提升信道状态信息(channel state information, CSI)指纹室内定位的性能,提出了一种改进MixNet的CSI图像指纹室内定位方法. 在离线阶段,通过选择定位参考点(reference point, RP)处信号强度指示(received signal strength indication, RSSI)最强的3个接入点(access point, AP),提取其CSI数据并转换为图像;然后利用改进的MixNet模型对图像进行训练并保存模型. 其中改进的MixNet引入了坐标注意力(coordinate attention, CA)和残差连接. 首先,将MixNet-s中的SE (squeeze-and-excitation)注意力替换为CA,以增强网络的信息表示能力并更精确地获取CSI图像指纹特征. 其次,根据MixNet-s模型的特点构建残差连接,以增强网络的表示能力并防止过拟合. 最后,通过减小网络深度确保所有网络层得到充分训练;在线阶段,采集目标设备的CSI数据并转换为图像,输入已训练好的改进MixNet模型(命名为MixNet-CA);最后利用加权质心算法根据模型输出的概率值估计目标设备的最终位置. 提出方法在室内环境中进行了验证,达到了0.3620 m的平均定位误差.

     

    Abstract: To enhance the performance of indoor localization using channel state information (CSI) fingerprints, an CSI image-based indoor localization method based on the improved MixNet model is proposed. In the offline phase, the method involves selecting the three access points (APs) with the highest received signal strength indication (RSSI) at the reference point (RP), extracting their CSI data, and converting it into image. Subsequently, the improved MixNet model is employed to train on these images and save the model. The improved MixNet model introduces coordinate attention (CA) and residual connections. Specifically, it replaces the squeeze-and-excitation (SE) attention in MixNet-s with CA to enhance the network’s information representation capability and extract CSI image fingerprint features more accurately. Moreover, it incorporates residual connections, tailored to the characteristics of the MixNet-s model, to enhance the network’s representation capacity and prevent overfitting. Finally, the network depth is reduced to ensure that all network layers are adequately trained. During the online phase, CSI data from the target device is collected and converted into image, and then input into the pre-trained improved MixNet model (named MixNet-CA). The final device position is estimated using a weighted centroid algorithm based on the model's output probabilities. The proposed method is validated in an indoor environment and achieve an average positioning error of 0.362 0 m.

     

/

返回文章
返回