基于RFID信号分析的非绑定式活动识别方法研究

发布时间：2018-06-17 22:15

本文选题：非绑定式活动识别 + 多普勒频移　；参考：《西北大学》2017年硕士论文

【摘要】：行为感知在健康看护、智能家居以及健身训练等应用中扮演着重要角色。其中,每年室内跌倒会造成世界上将近40万的死亡人数,因而室内活动识别逐渐受到学术界和工业界的密切关注。近几年来,绑定式和非绑定式活动识别技术层出不穷。绑定式活动识别研究要求用户时时刻刻携带可穿戴设备才能检测日常行为,使得无需用户携带任何通信设备的非绑定式活动识别被广泛采纳和认可。目前具有普适性的非绑定式活动识别在实际应用中也面临着诸多挑战:不同室内监测环境存在不同信号反射路径而导致活动指纹无法适用于不同监测环境,不同活动执行的速度和方向会降低活动匹配准确度,等等。针对以上两大问题,本文研究基于RFID信号分析的非绑定式活动识别方法,设计一种利用低成本设备实现能应用于多个室内监测环境的高精度活动识别方法。本文的主要研究内容和创新点包括以下两个方面:(1)从低成本、低采样率设备中提取对多径条件具有鲁棒性的活动指纹。利用MUSIC算法估计空间谱的思想,设计出一种时域上的频谱估计方法,提取人活动引起的无线信号收发设备间通信路径长度变化速度,即多普勒频移,建立对多径变化不敏感的多普勒频移指纹,有效避免了为保持活动识别精度而针对每个监测环境采集活动指纹,降低人力代价。(2)设计活动匹配算法,降低活动速度和方向多样性带来的匹配误差。通过实验分析,活动执行速度和方向多样性会引起多普勒频移在时间长度和幅度上的波动,本文采用DTW算法解决时间长度上的波动带来的匹配误差,又进一步改进DTW算法,动态调整两个指纹波峰波谷之差所占两条指纹差距的比例,降低多普勒频移指纹在幅度上波动带来的匹配误差。从而在训练一次的前提下,准确识别执行速度和方向多样性的活动。最后,本文利用商业RFID系统,在三种不同典型的室内场景(实验室、家庭场所、图书馆)下进行真实实验,实验结果显示本文方法的活动识别准确率分别为92%、93%、91%,从而验证了本文方法能在训练一次的情况下实现不同监测环境中的活动识别。
[Abstract]:Behavioral perception plays an important role in health care, smart home and fitness training. Among them, indoor falls cause nearly 400000 deaths in the world every year, so indoor activity identification has been paid close attention to by academia and industry. In recent years, binding and non-binding activity recognition techniques have emerged in endlessly. Binding activity recognition requires users to carry wearable devices all the time in order to detect daily behavior, which makes unbound activity recognition without any communication devices widely adopted and recognized. At present, there are many challenges in the practical application of universal unbound activity recognition: different signal reflection paths exist in different indoor monitoring environments, which lead to the fact that the active fingerprints can not be applied to different monitoring environments. Different activity execution speed and direction will reduce the accuracy of activity matching, and so on. Aiming at the above two problems, this paper studies the unbound activity identification method based on RFID signal analysis, and designs a high-precision activity identification method which can be used in multiple indoor monitoring environments by using low-cost equipment. The main research contents and innovations in this paper include the following two aspects: 1) extracting active fingerprints which are robust to multipath conditions from low cost and low sampling rate devices. Using music algorithm to estimate spatial spectrum, a time-domain spectrum estimation method is designed to extract the varying speed of communication path length between wireless signal transceiver devices caused by human activities, that is, Doppler frequency shift. The Doppler frequency shift fingerprint, which is insensitive to multipath change, is established, which effectively avoids the design of activity matching algorithm for each monitoring environment in order to maintain the accuracy of activity identification and reduce the human cost. Reduce the matching error caused by velocity and direction diversity. Through experimental analysis, the diversity of execution speed and direction will cause the wave of Doppler frequency shift in time length and amplitude. In this paper, DTW algorithm is used to solve the matching error caused by the fluctuation of time length, and the DTW algorithm is further improved. Dynamically adjust the proportion of the difference between two fingerprint peaks and troughs in order to reduce the matching error caused by the fluctuation of Doppler frequency shift fingerprint. In order to accurately identify the execution speed and direction of the diversity of activities under the premise of one training. Finally, this paper uses the commercial RFID system to carry out real experiments under three different typical indoor scenes (laboratory, home, library). The experimental results show that the accuracy of activity recognition of this method is 922 / 931, respectively, which verifies that the method in this paper can realize activity recognition in different monitoring environments under the condition of training once.
【学位授予单位】：西北大学
【学位级别】：硕士
【学位授予年份】：2017
【分类号】：TP391.44

【相似文献】