Abstract:

With the widespread use of mobile phone sensors, there are more and more demands for daily human behavioral action recognition. Classical recognition methods use heuristic processes to obtain artificial features, and then use machine learning methods to identify actions. The latest research shows that the Inception convolution structure is prominent, which can avoid the deviation problem caused by the manual extraction of features. Human action is composed of complex motion sequences. It is essential for action recognition to catch this time series.Based on this, this paper first optimizes the Inception structure, proposes the O-Inception structure, and fuses it with the long short term memory model (LSTM), last proposes the optimization Inception-LSTM (OI-LSTM) action recognition model. On the one hand, OI-LSTM model can use O-Inception structure to extract features automatically. On the other hand, it can also use LSTM to capture the action sequence, and then improve the accuracy of action recognition. Experiments are conducted on the wireless sensor data mining (WISDM) and UC Irvine (UCI) datasets. The experimental results show that the proposed OI-LSTM action recognition model has an accuracy improvement of about 4% on the WISDM dataset and 1% on the UCI compared to the current state-of-the-art method. Experi-ments also show that the model has strong fault tolerance and real-time performance.

Key words: mobile phone sensor analysis, human activity recognition, Inception, convolutional neural network (CNN), depth-wise separable convolution, long short term memory (LSTM), deep learning, optimization Inception-LSTM (OI-LSTM)

摘要：

随着手机传感器的普遍使用，对人体日常行为动作识别需求越来越多，经典识别方法是利用启发式过程获得人工特征，再用机器学习方法识别动作。最新研究表明，Inception卷积结构在特征自动提取方面表现尤为突出，可避免人工提取特征带来的偏差问题。人体动作由复杂运动序列构成，捕捉该时间序列是动作识别必不可少的。基于此，首先对Inception结构进行了优化，提出了O-Inception结构，并将其与长短期记忆模型（long short term memory，LSTM）进行了融合，进而提出了OI-LSTM（optimization Inception-LSTM）动作识别模型。OI-LSTM模型一方面可以利用O-Inception结构实现对特征的自动提取，另一方面，还可以利用LSTM捕获动作时序，进而提高了动作识别准确率。在WISDM（wireless sensor data mining）和UCI（UC Irvine）两个数据集上进行了扩展性实验，实验结果表明，所提出的OI-LSTM动作识别模型，在WISDM和UCI两个数据集上其准确率比当前最先进的方法分别提高了约4%和1%。实验还证明，此模型拥有很强的容错性和实时性。

关键词: 手机传感器分析, 人体动作识别, Inception, 卷积神经网络（CNN）, 深度可分离卷积, 长短期记忆模型（LSTM）, 深度学习, OI-LSTM