This paper proposes a d-vector approach for extracting robust biometrics from inertial signals recorded with wearable sensors. The d-vector approach generates identity representations using a deep learning architecture composed of Convolutional Neural Networks. This architecture includes two convolutional layers for learning features from the inertial signal spectrum. These layers were pretrained using data from 154 subjects. After that, additional fully connected layers were attached to perform user identification and verification, considering 36 new subjects. This paper compares the proposed d-vector approach with previous proposed algorithms using in-the-wild recordings in different scenarios. The results demonstrated the robustness of the proposed d-vector approach for in-the-wild conditions: 97.69% and 94.16% accuracies (for user identification) and 99.89% and 99.67% Areas Under the Curve (for user verification) were obtained using one (walking) or several activities (walking, jogging and stairs) respectively. These results were also verified in laboratory conditions improving the performance reported in previous works. All the analyses were carried out using public datasets recorded at the Wireless Sensor Data Mining laboratory.

本文提出了一种d矢量方法，用于从可穿戴式传感器记录的惯性信号中提取鲁棒的生物特征。d向量方法使用由卷积神经网络组成的深度学习架构来生成身份表示。该架构包括两个卷积层，用于从惯性信号频谱中学习特征。使用来自154位受试者的数据对这些图层进行了预训练。此后，考虑了36个新主题，附加了附加的完全连接的层以执行用户标识和验证。本文将提出的d矢量方法与先前提出的在不同情况下使用实时录制的算法进行比较。结果证明了提出的d矢量方法在野外条件下的鲁棒性：97.69％和94。分别使用一项（步行）或多项活动（步行，慢跑和上楼梯）获得了16％的准确度（用于用户识别）以及99.89％和99.67％的曲线下面积（用于用户验证）。这些结果还在实验室条件下得到了验证，从而改善了先前工作中报告的性能。所有分析均使用无线传感器数据挖掘实验室记录的公共数据集进行。