Human Activity Recognition (HAR) has many important applications in health care which include management of chronic conditions and patient rehabilitation. An important consideration when training HAR models is whether to use training data from a general population (subject-independent), or personalised training data from the target user (subject-dependent). Previous evaluations have shown personalised training to be more accurate because of the ability of resulting models to better capture individual users' activity patterns. However, collecting su cient training data from end users may not be feasible for real-world applications. In this paper, we introduce a novel approach to personalised HAR using a neural network architecture called a matching network. Matching networks perform nearest-neighbour classi cation by reusing the class label of the most similar instances in a provided support set. Evaluations show our approach to substantially out perform general subject-independent models by more than 5% macro-averaged F1 score.
Automatic recognition and tracking of human activity using wearable sensors
is increasingly being adopted for health care applications e.g. management of
chronic low back pain in SelfBACK 1 [
1The SelfBACK project is funded by European Union's H2020 research and innovation programme under grant agreement No. 689043.
In this paper, we introduce an approach to personalised HAR using matching
networks. Matching Networks are a type of neural network architecture
introduced for the task of one-shot learning [
The aim of matching networks is to learn a model that maps an unlabelled example x^ to a class label y^ using a small support set S of labelled examples. This is illustrated in Figure 1. Given a set of instances X = fxg and a set of class labels L = fyg, an embedding function f which in this case a neural network parameterised by , the function a is an attention mechanism that takes the embedded representation of a test instance x^ and a support set S and returns a probability distribution P (yjx^; S) over class labels y of instances in S. To train the matching network for personalised HAR, we also de ne a set of users U where each user uj 2 U is comprised of a set of labelled examples as follows: Next we de ne a set of training instances Tj for each user uj as follows:
Tj = f(Sj ; Bj )gl i.e., Tj is made up of user-speci c support and target set pairs Sj and Bj respectively, where Sj = f(x; y)jx 2 ui; y 2 Lg and Bj = f(x; y)jx 2 uj ; x 62 Sj g. Note that the set of labels in Sj is always equivalent to L because we are interested in learning a classi er over the entire set of activity labels. Accordingly, Sj contains m examples for each class y 2 L and the cardinality of Sj is jSj j = m jLj. Both Sj and Bj are sampled at random from uj l times to create Tj . Each Bj is used with it's respective Sj by classifying each instance in Bj using Sj and computing loss using categorical cross entropy. The network is trained using stochastic gradient descent and back propagation. (1) (2) 3
Evaluation is conducted on a dataset of 50 users with 9 activity classes (Standing, Sitting, Lying, Walking Slow, Walking Normal, Walking Fast, Up Stairs, Down Stairs) where each user performs each activity for about 3 minutes. We adopt a hold-out validation strategy where 8 out of the 50 users are randomly selected for testing. To simulate user provided samples for creating personalised support sets, we hold out the rst 30 seconds of each test user's data for creating the support set. This leaves approximately 150 seconds of data per activity which are used for testing, Performance is reported using macro-averaged F1 score.
In the evaluation, we explore the performance of our matching network against a number of baseline approaches. Accordingly we compare the following algorithms: { kNN: Nearest-neighbour classi er trained on the entire training set { SVM: Support Vector Machines trained on the entire training set { MLP: A Feed-forward neural network trained on the entire training { MNet: Our personalised matching network approach All algorithms use 5-second window sizes for signal partitioning and discrete cosine transform for feature representation. performs both SVM and MLP by more than 5% which shows the e ectiveness of our matching network approach at exploiting personal data for activity recognition. 4
In this paper, we presented a novel approach for personalised HAR using matching networks. There are two main advantages to the approach we presented. Firstly, our approach is able to achieve high accuracy using only a small set of user provided examples (30 seconds in this work) which makes it more practical for real-world applications compared to subject-dependent training which requires the end user to provide large amounts (possible hours) of labelled training data. Secondly, our approach does not require retraining the model at runtime when new data becomes available which makes it very adaptable.