在基于RFID的应用中，获取细粒度的空间信息具有重要的实际意义。 然而，在现有的商用（COTS）RFID系统中，高精度定位仍然是一项具有挑战性的任务。 受计算机视觉（CV）领域进展的启发，研究人员提出将CV与RFID系统相结合，将定位问题转化为匹配问题。 目前的方法通过将CV提取的轨迹转换为相位序列进行匹配，从而实现CV与RFID的融合，这是一种降低空间分辨率的方法。 因此，它们在更恶劣的条件下无法正常工作，例如标签间隔小和标签读取率低。 为了解决这一局限性，我们提出了TagFocus，这是一个更强大的支持RFID的系统，它利用视觉辅助信息进行细粒度多对象识别和跟踪。 TagFocus的关键观察是，如果从同一个对象获取不同方法生成的轨迹，则这些轨迹应是兼容的。 利用这一观察，系统训练了一个基于注意力机制的序列到序列（seq2seq）模型，它为每个候选目标对生成一条预测运动轨迹。 而正确配对的轨迹应与CV直接提取的观测轨迹最为匹配。 TagFocus的原型在实验室环境中实现并进行了全面评估。 实验结果表明，TagFocus算法在匹配精度和鲁棒性方面均优于现有的算法。
In recent years, with the rapid development of industrial Internet of things, the mar-ket demand for smart logistics, smart medical and so on has surged. As an important partof Internet of things, radio frequency identification (RFID) technology has been widelyconcerned by industry and academia. Among them, RFID based positioning technologyhas become one of the indoor positioning technologies that researchers focus on becauseof its low cost, strong universality and easy deployment. How to improve the position-ing accuracy and stability of RFID based positioning technology has become an urgentproblem to be solved in large-scale application of the technology.Most of the existing RFID based positioning methods collect the received signalstrength (RSSI), phase and other wireless signal characteristics of the tag through RFIDreader, and directly calculate the location of the tag after analyzing the above charac-teristics. However, due to the inherent hardware characteristics of the device, multipathinterference in the environment and other factors, there will inevitably be a certain loss ofaccuracy when calculating the label position, which will further increase the positioningerror.In this paper, a multi-target location and tracking algorithm based on the fusion ofRFIDandcomputervision(CV)isproposedtomeettheneedsofbloodsamplelocationinmedical industry. The algorithm makes full use of the wireless signal characteristics col-lectedbyRFIDreaderinthesamplemanagementplatformandthemovingvideorecordedby monocular camera in the process of sample insertion / removal, and realizes the accu-rate positioning and tracking of multiple samples. The core idea of the algorithm is thatthere is a certain non-linear correlation between the change of wireless signal character-istics of the sample label and the movement trajectory of the sample. We use a sequenceto sequence (seq2seq) model based on attention mechanism to learn it, so as to completethe conversion process from RFID signal features to motion trajectory.Theexperimentalresultsshowthatthemulti-targetlocationandtrackingtechnologybased on RFID and CV can achieve more than 96% accuracy in general scenes, whichproves that the system has good feasibility and robustness in real application scenarios.
