Weakly Supervised Human Activity Recognition From Wearable Sensors By Recurrent Attention Learning

Human activity recognition is a research hotspot in the field of artificial intelligence and pattern recognition.With the continuous advancement of smart wearable device and deep learning,human activity recognition based on wearable sensors has become an important research direction.Although these traditional human activity recognition method based on deep learning can automatically learn the appropriate classification features without relying on the researcher's experience,they still belong to the scope of supervised learning and require good training datasets in which thousands of training sequences should be carefully labeled.However,unlike images or videos which can be easily classified by human beings,strictly labeling such sequences of sensor data needs much more manpower and computing resources.Therefore,whether the weakly labeled dataset can be used to train the model and how to design a deep learning model for weakly labeled sensor data has become the main issue for consideration.In view of the above problems,we present a new weakly supervised human activity recognition model based on recurrent attention learning,in which an agent is trained to extract information from weakly labeled sensor data by adaptively selecting a sequence of locations,and the following aspects are completed:(1)In this paper,the accelerometer sensor in the smart phone is used to collect the acceleration data generated during the human activity.And then,the data is preprocessed by the general windowing and standardization operations to generate the target weakly labeled sensor dataset.(2)For the traditional activity recognition model,the entire sensor data information needs to be used,and it is only applicable to the short strongly labeled sensor data.In this paper,we present a new weakly supervised human activity recognition model from wearable sensors by recurrent attention learning,which uses convolutional neural network(CNN)and recurrent neural network(LSTM)to construct a deep learning network.In order to extract the high-value feature from input data quickly,attention mechanism and reinforcement learning are introduced.By mimicking human visual attention,the model can quickly find the most representative part of the data within a small number of steps to improve the accuracy of the classification.(3)Since the model is non-differentiable,it is trained by reinforcement learning.Because multiple activities may occur in a sequence of sensor data,we design two novel reward strategies for two different weakly labeled datasets to achieve accurate positioning of multiple activities in the data.(4)In order to evaluate the weakly supervised model proposed in this paper,two baseline models are constructed in this paper and the three models were tested on the weakly labeled dataset and the traditional strongly labeled dataset.The experimental results show that the weakly supervised model proposed in this paper is superior to the traditional deep learning model on both datasets.