Device-free localization (DFL) systems have aroused extensive attention because it is more convenient than device-enabled localization systems, and fingerprint-based localization method is usually used in DFL systems. Although fine-grained information can be provided by the channel state information (CSI), but changes in the environment over time can cause the CSI become different. Therefore, the real-time CSI data can’t match with the data in the fingerprint map established beforehand very well, which can lead to the inaccuracy of the positioning result. This paper presents a DFL system, which adopts transfer learning method to update the fingerprint map and employs the Light Gradient Boosting Machine (LightGBM) algorithm to train the fingerprint map. Wavelet transform is used in this paper to filter the noise in the raw CSI data and the CSI data on a portion of the fingerprint points are collected to update the established fingerprint map by joint distribution adaptation in the update stage. After classifying the CSI data of the testing point by LightGBM, the position coordinate is achieved by the confidence regression method. By using LightGBM, the proposed system can achieve the average distance error of 0.48m, outperforming the result by using eXtreme Gradient Boosting (XGBoost) and Gradient Boost Decision Tree (GBDT). According to the result of the four-week experiment, the average distance error of this system can be decreased by 21% compared with not using the calibration method.

无设备定位（DFL）系统引起了广泛关注，因为它比启用设备的定位系统更方便，并且基于指纹的定位方法通常用于DFL系统中。尽管可以通过信道状态信息（CSI）提供细粒度的信息，但是环境随时间的变化会导致CSI变得不同。因此，实时CSI数据与预先建立的指纹图中的数据不能很好地匹配，从而导致定位结果的准确性。本文提出了一种DFL系统，该系统采用转移学习的方法更新指纹图谱，并采用光梯度增强机（LightGBM）算法训练指纹图谱。本文采用小波变换对原始CSI数据中的噪声进行滤波，并收集部分指纹点的CSI数据，通过在更新阶段进行联合分布自适应来更新已建立的指纹图谱。用LightGBM对测试点的CSI数据进行分类后，通过置信度回归方法获得位置坐标。通过使用LightGBM，提出的系统可以实现0.48m的平均距离误差，优于使用极限梯度增强（XGBoost）和梯度增强决策树（GBDT）的结果。根据四周实验的结果，与不使用校准方法相比，该系统的平均距离误差可以减少21％。