3.2.3. Machine Learning Methods

Machine learning techniques such as supervised learning, unsupervised learning, online learning and transfer learning have been investigated for activity recognition and walk detection.

Supervised learning such as decision trees (DT) [56], neural networks (NN), support vector machines (SVM), Gaussian mixture models (GMMs) [32], k-nearest neighbor (KNN) [29], naive Bayes classifiers [28] and boosting methods [57] have been studied and generally achieve good detection accuracy.

Unsupervised methods have been used in activity recognition. For example, hidden Markov models [36,58–60] are powerful in sequence data analysis; they could also be used to model the walking activity. K-means clustering [61,62] clusters the data in feature space, where the activities could be identified. Although the recognition accuracy of unsupervised learning is generally lower than supervised learning [36], it exempts people from the costly work of labeling the training data.

Besides these typical learning approaches, Cheng et al. [63] investigated zero-shot learning that could recognize unseen new activities when there were no corresponding samples in the training dataset. Rebetez et al. [64] introduced growing neural gas (GNG) to build an online learning recognition system that did not require labeled data. Transfer learning [65,66] could transfer activity recognition from one domain to another domain, which adapts the changes of sensor position [33,67], activity type [68] or environment scenario [69].
