南静, 建中华, 宁传峰, 代伟. 一种轻量型人体行为识别学习模型[J]. 工程科学学报, 2022, 44(6): 1072-1079. DOI: 10.13374/j.issn2095-9389.2021.03.18.001
引用本文: 南静, 建中华, 宁传峰, 代伟. 一种轻量型人体行为识别学习模型[J]. 工程科学学报, 2022, 44(6): 1072-1079. DOI: 10.13374/j.issn2095-9389.2021.03.18.001
NAN Jing, JIAN Zhong-hua, NING Chuan-feng, DAI Wei. Lightweight human activity recognition learning model[J]. Chinese Journal of Engineering, 2022, 44(6): 1072-1079. DOI: 10.13374/j.issn2095-9389.2021.03.18.001
Citation: NAN Jing, JIAN Zhong-hua, NING Chuan-feng, DAI Wei. Lightweight human activity recognition learning model[J]. Chinese Journal of Engineering, 2022, 44(6): 1072-1079. DOI: 10.13374/j.issn2095-9389.2021.03.18.001

一种轻量型人体行为识别学习模型

Lightweight human activity recognition learning model

  • 摘要: 提出一种基于近邻成分分析(Neighbourhood component analysis, NCA)、L2正则化和随机配置网络(Stochastic configuration networks, SCNs)的轻量型人体行为识别学习模型. 首先, 针对人体行为特征集维数过高且可分性差的问题, 利用NCA从特征集中选择高相关性特征子集, 进而提高模型建模计算过程的轻量性和识别精度. 其次, 针对SCNs隐含层节点过多时容易出现过拟合的问题, 采用L2正则化方法增强SCNs的泛化能力, 同时利用监督机制约束产生隐含层参数的方法, 极大地提高了SCNs模型的轻量性. 最后, 将所提NCA−L2−SCNs学习模型在UCI HAR特征集上进行验证, 实验结果表明, 相比于其他模型, 本文所提轻量型模型对于人体行为识别具有更好的识别精度和更快的建模速度.

     

    Abstract: In the past few decades, smartphone-based human activity recognition research has played an important role in many fields, including smart buildings, healthcare, and the military. However, the CPU and storage space of smartphones are very limited, so developing a lightweight human activity recognition learning model has become a research focus and hot spot in this field. To address the abovementioned problems, this paper proposed a lightweight human activity recognition learning model based on the nearest neighbor component analysis (NCA), L2 regularization, and stochastic configuration networks (SCNs). In the proposed model, aiming first at the problem of high dimension and poor separability exhibited by the human activity data, NCA was used to select a subset of highly relevant data from the dataset to improve the lightness of calculation using the learning algorithm in the modeling process and recognition accuracy of the established model. Second, to prevent the occurrence of overfitting when there are too many hidden layer nodes in SCNs, the L2 regularization method was adopted to enhance the generalization ability of SCNs. At the same time, the method of using the supervision mechanism to restrict the generation of hidden layer parameters greatly improved the lightness of the SCNs model. Finally, the proposed learning model and other learning models were verified experimentally on the UCI human activity recognition dataset. Experimental results show that compared with SCNs, the proposed L2−SCNs model reduces the lightness of the number of parameters by 20% and helps improve the accuracy of the model. The introduction of the NCA method has greatly facilitated the recognition accuracy and lightness (modeling time) of the L2−SCNs model, increasing by 3.41% and 70.24%, respectively. Moreover, compared with other state-of-the-art models, such as the support vector machine and long short-term memory network, the proposed model achieves the best recognition accuracy of 97.48% in the shortest time. To sum up, the model proposed herein is a lightweight human activity recognition model with exceptional recognition accuracy and a fast modeling speed.

     

/

返回文章
返回