Improved Human Activity Recognition Model based on Sequence Mixup
Ran Tan, Xuefeng Yan · 2020
In the practical application of smartphone-based human activity recognition (HAR), the location and orientation of smartphones are diverse and uncertain, which requires high robustness of deep learning model. Mixup is a novel regularization method, which has been successful in the field of image recognition, but it has the problem of amplitude information loss when applied to sensor data. The sequence mixup (S-mixup) is proposed, which replaces linear mixing in amplitude with interpolation in the time dimension. The proposed method can avoid the problem of amplitude information loss, stimulate the model to summarize the law of amplitude and orientation change, and extract features unrelated to smartphone orientation, to prevent overfitting. The experimental results show that the improved deep learning model based on the proposed method has high recognition accuracy when dealing with the adversarial examples with different smartphone orientations, which proves the effectiveness of the proposed method.