基于RFID的轻量化的动作识别方法

RFID-based lightweight action recognition method

针对基于射频识别(RFID)的相位(Phase)信号的动作识别技术识别精度不高或不够轻量化等问题,提出基于RFID的轻量化的动作识别方法。该方法通过格拉姆矩阵将一维数据转化为二维图像,将Phase转换为格拉姆角场作为改进的MobileNet网络输入,通过格拉姆角场图像所拥有的高阶信号描述能力表现更多的几何性质和内在数据结构,在减小异常数据对识别效果影响的同时提高识别性能并充分利用网络。同时,将坐标注意力机制融合压缩激励机制,获取相位信号中更丰富的上下文信息使模型更好地定位和识别目标,同时在保证不影响识别精度的情况下重新构建网络,减少模型参数以及所需计算量。该方法与传统MobileNet相比,模型参数量仅为原模型的12.9%,同时识别率提高2.34%,每秒浮点运算次数也优于原模型。实验结果表明,该模型各个指标都表现出更优的实验结果,能够完成动作识别的相关要求。

In view of the low accuracy and the lack of lightweight of action recognition technology based on radio frequency identification(RFID)phase signals,a lightweight RFID-based action recognition method is proposed.The one-dimensional data is transformed into two-dimensional images by the Gram matrix.The phase signals are converted into the Gram angular field,which is taken as the input of the improved MobileNet.More geometric properties and intrinsic data structures of the image is exhibited by the Gramian angular field(GAF)images′higher-order signal description capabilities,so as to improve the recognition performance of the method and utilize the network fully while reducing the influence of abnormal data on the recognition effect.The coordinate attention(CA)mechanism is integrated with the squeeze-and-excitation mechanism,so that the model can capture more abundant contextual information from the phase signals and obtain better object localization and recognition.The network is reconstructed to reduce the model parameters and the required computational load,without affecting its recognition accuracy.In comparison with the traditional MobileNet,the quantity of the parameter of the proposed model are only 12.9%of the original model,its recognition rate is improved by 2.34%,and its floating-point operations per second(FLOPs)is also superior to that of the original model.The experimental results show that the proposed model exhibits superior performance for all indexes and can fulfill the requirements for action recognition tasks.