Lifelong Learning in Sensor-Based Human Activity Recognition

Abstract

Human activity recognition systems will be increasingly deployed in real-world environments and for long periods of time. This significantly challenges current approaches to human activity recognition, which have to account for changes in activity routines, evolution of situations, and of sensing technologies. Driven by these challenges, in this article we argue the need to move beyond learning to lifelong machine learning-with the ability to incrementally and continuously adapt to changes in the environment being learned. We introduce a conceptual framework for lifelong machine learning to structure various relevant proposals in the area, and identify some key research challenges that remain. A key enabler for many critical ubiquitous computing applications, in areas such as healthcare (e.g., mental well-being assessment, and clinical assessment on cognition and mobility) and home automation (e.g., automatic heating configurations) is sensor-based human activity recognition (HAR): the ability automatically to recognise and predict users' current and future activities from data collected from a wide range of wearable and ambient sensors. Based on inferred user activities, applications can deliver customised, in situ services in an automatic and unobtrusive manner. Activity recognition has been extensively studied for over 10 years, and significant progress has been made through the use of modern data-driven techniques, including Hidden Markov Models, Conditional Random Fields, Support Vector Machines, and-more recently-deep neural networks. These techniques have achieved promising results in learning complex correlations between sensor data and activities of interest: their success in recognising human activities in lab and testbed settings is now enabling the move toward large-scale, in-the-wild, and long-term deployment of activity recognition systems [17]. This move however comes with its own chal

XLearn

“…Also to enable activity space remapping, we need to take a collection of pre-defined activities from each dataset to compute the activity similarity matrix. With a general methodology of human activity recognition [40], the system designer pre-defines a closed set of activities of interest, which are often the requirements from applications such as personal healthcare. Then driven by the activities, the designer will select a range of ambient and/or wearable sensors that can potentially detect these activities.…”

Semi-Supervised Adversarial Auto-Encoder to Expedite Human Activity Recognition

“…The model must be rebuilt and retrained when adding a new activity class. Few studies have looked into the potential for an activity model to emerge automatically with various activities [ 17 ]. However, this capability has the advantage of maintaining the knowledge in the time-tested business model while reducing the need for manual feature engineering, manual configuration, and training expenses.…”

A multi-scale feature extraction fusion model for human activity recognition

“…Different data patterns represent different ways of encoding human behavior, providing different values and sources of information. Currently, HAR can be broadly classified into two categories based on the source of data acquisition: video-based 5 – 7 and sensor-based 8 – 10 . Video-based systems mainly use devices such as cameras to capture videos and images to recognize daily life activities and human behaviors through techniques in computer vision.…”