Learning Deep and Shallow Features for Human Activity Recognition

Abstract

selfBACK is an mHealth decision support system used by patients for the self-management of Lower Back Pain. It uses Human Activity Recognition from wearable sensors to monitor user activity in order to measure their adherence to prescribed physical activity plans. Different feature representation approaches have been proposed for Human Activity Recognition, including shallow, such as with hand-crafted time domain features and frequency transformation features; or, more recently, deep with Convolutional Neural Net approaches. The different approaches have produced mixed results in previous work and a clear winner has not been identified. This is especially the case for wrist mounted accelerometer sensors which are more susceptible to random noise compared to data from sensors mounted at other body locations e.g. thigh, waist or lower back. In this paper, we compare 7 different feature representation approaches on accelerometer data collected from both the wrist and the thigh. In particular, we evaluate a Convolutional Neural Net hybrid approach that has been shown to be effective on image retrieval but not previously applied to Human Activity Recognition. Results show the hybrid approach is effective, producing the best results compared to both hand-crafted and frequency domain feature representations by a margin of over 1.4% on the wrist.

Improving kNN for Human Activity Recognition with Privileged Learning Using Translation Models

“…We use DCT feature transformations as it has been proven to result in significant performance improvements over raw multi-dimensional features. DCTs extract generic and robust features compared to other statistically crafted features and was also shown to have slightly better or comparable results to deep feature embeddings [9]. Importantly for us, it simplifies the task of translators when the mapping can be carried over a proven feature representation for input and output data.…”

“…Typically these applications relate to tracking or monitoring movements such as ambulatory activities (i.e. running or jogging) [9,11], daily activities of living (i.e. gardening or cooking) [1] or exercises (i.e.…”

“…Much of the latest research has focussed on exploiting multiple sensors for HAR with deep learning models to achieve state of the art performance [16,8]. In [9] the authors explore the impact of different sensor placements on HAR performance and discuss the tradeoff between convenience (wrist placement) versus accuracy (thigh placement). Ideally we want to optimise sensor placement convenience whilst minimising the negative impact this can have on accuracy.…”

Personalised Human Activity Recognition Using Matching Networks

“…Many different feature extraction approaches have been applied for HAR. These include hand-crafted time and frequency domain features, coefficients of frequency domain transformations, as well as more recent deep learning approaches [17]. One feature extraction approach we have previously found to be both inexpensive to compute and very effective, is Discrete Cosine Transform (DCT) [17].…”

“…These include hand-crafted time and frequency domain features, coefficients of frequency domain transformations, as well as more recent deep learning approaches [17]. One feature extraction approach we have previously found to be both inexpensive to compute and very effective, is Discrete Cosine Transform (DCT) [17]. DCT is applied to each axis (a i , b i , c i ) of a given window w i to produce vectors of coefficients v a , v b and v c respectively that describe the sinusoidal wave forms that constitute the original signal.…”