Research On Human Activity Recognition For Wearable Devices Based On Deep Domain Adaptation

With the continuous expansion of the wearable device market,smart watches and wristbands are seamlessly connected to people's daily life.These devices have spawned a variety of activity recognition applications with their various motion sensors,such as motion type recognition and micro-gesture recognition,these applications play an important role in many human computer interaction(HCI)scenarios.We mainly focus on human activity recognition based on wearable devices.We first take motion type recognition as the entry point to propose and implement a variety of solutions,including manually extracting statistical time/frequency domain features and combining them with proper classifiers,converting the original data into a two-dimensional time-frequency spectrogram by Short-Time Fourier Transform(STFT)and inputting it into Convolutional Neural Network(CNN)for training and recognition.At the same time,corresponding algorithms(Borderline-SMOTE1 and post-processing algorithm based on the sliding window)are designed to improve the recognition effect for class imbalance and data jump problems in the process of data collection.Considering the distribution discrepancy of the data collected by different users and different devices,the above methods cannot achieve good results in the scene of cross-user and cross-device activity recognition.Inspired by domain adaptation strategies,we proposed XHAR,an adversarial deep domain adaptation framework for adaptation between different users and different smart devices.XHAR first selects a source data set(with label)which is the most similar to the target data set(without label),and then utilizes CNN and Bidirectional Gated Recurrent Unit(BiGRU)to extract features,at the same time,we design attention mechanism for the combination of different features.We establish multiple domain discriminators to reduce the distribution discrepancy and finally perform adaptation on the target data set to obtain the predicted labels.We conduct extensive experiments on 50 users and 4 smart devices with two kinds of datasets,i.e.,micro-gesture data set and motion-type data set.We compare XHAR with the source-only model and several state-of-the-art domain adaptation models.The experimental results show that the classification accuracy of XHAR improves at least 4.81%(to 74.50%)between different users and 9.25%(to 69.23%)between different devices.In terms of real-time recognition,the average recognition time for a single action is 6.27545 ms.