Similarity Embedding Networks for Robust Human Activity Recognition

Deep learning models for human activity recognition (HAR) based on sensor data have been heavily studied recently. However, the generalization ability of deep complex real-world HAR is limited by availability high-quality labeled data, which are hard to obtain. In this article, we design a similarity embedding neural network that maps input signals onto real vectors through carefully designed convolutional and Long Short-Term Memory (LSTM) layers. The trained with pairwise loss, encouraging clustering samples from same class in embedded space, can be effectively small dataset even noisy mislabeled samples. Based learned embeddings, further propose both nonparametric parametric approaches recognition. Extensive evaluation two public datasets has shown proposed significantly outperforms state-of-the-art classification tasks, robust training set, also used denoise dataset.