Object recognition, localization and private object interaction tracking for maintenance-free IoT

对象识别：我们准备并向IEEE IoT Journal提交了一篇完整的论文。我们开发了RadioRec，这是一个基于深度学习的系统，可以根据日常物品与微波信号的交互以非接触方式识别它们。与我们早期的研讨会论文相比，RadioRec在学习能力和评估方面都有了重大升级。RadioRec的工作原理是使用单个天线对通过物体传输微波信号。RadioRec使用自动编码器自动提取特征，并使用它们来训练对象分类模型。我们的评估表明，RadioRec可以检测和识别26种不同材料和形状的日常物体，准确率超过97%。审查人员要求在不同的环境下进行额外评价，并在案文中作出若干澄清。我们现在正在处理审稿人的意见，并按照审稿人的建议，准备向同一期刊提交最新的投稿。到目前为止，我们已经在另一个环境中进行了额外的实验，具有各种对象方向和对象位置，并获得了有希望的初步结果。（DOI：10.1145/2785956.2787487），适用于USRP软件-定义的无线电平台，以实现对后向散射标签上的到达角和飞行时间的同时估计。我们目前正致力于在受控环境中正确进行估计（目前，没有反向散射标签）。

Object recognition: We prepared and submitted a full paper to the IEEE IoT Journal. We developed RadioRec, a deep learning-based system that recognizes everyday objects based on their interactions with microwave signals in a contact-less manner. RadioRec is a significant upgrade compared to our earlier workshop paper, both in terms of the learning capabilities and in terms of evaluation. RadioRec works by transmitting a microwave signal through the object using a single antenna pair. RadioRec extracts features automatically using an autoencoder, and uses them to train an object classification model. Our evaluation shows that RadioRec can detect and recognize 26 everyday objects of various materials and shapes with an accuracy of over 97%.The paper was rejected. The reviewers requested additional evaluation in a different environment, and also several clarifications in the text. We are now working on addressing the reviewers’ comments and preparing an updated submission to the same journal, as suggested by the reviewers. So far, we have performed additional experiments in another environment, with various object orientations and object positions, and have obtained promising preliminary results.Object localization: We have prepared a preliminary implementation of SpotFi (DOI: 10.1145/2785956.2787487) for the USRP software-defined radio platform to enable the simultaneous estimation of angle of arrival and time of flight on backscatter tags. We are currently working on getting the estimation to work correctly in a controlled environment (for now, without backscatter tags).