This paper reviews the research literature on human motion analysis using inertial sensors with the aim to find out: which configuration of sensors have been used to measure human motion; which algorithms have been implemented to estimate position and orientation of segments and joints of human body; how the performance of the proposed systems has been evaluated; and what is the target population with which the proposed systems have been assessed. These questions were used to revise the current state-of-the-art and suggest future directions in the development of systems to estimate human motion. A search of literature was conducted on eight Internet databases and includes medical literature: PubMed and ScienceDirect; technical literature: IEEE Xplore and ACM Digital Library; and all-science literature: Scopus, Web of Science, Taylor and Francis Online, and Wiley Online Library. A total of 880 studies were reviewed based on the criteria for inclusion/exclusion. After the screening and full review stages, 37 papers were selected for the review analysis. According to the review analysis, most studies focus on calculating the orientation or position of certain joints of the human body, such as elbow or knee. There are only three works that estimate position or orientation of both, upper and lower limbs simultaneously. Regarding the configuration of the experiments, the mean age of the test subjects is 26.2 years (± 3.7), indicating a clear trend to test the systems and methods using mainly young people. Other population groups, such as people with mobility problems, have not been considered in tests so far. Human motion analysis is relevant for obtaining a quantitative assessment of motion parameters of people. This assessment is crucial for, among others, healthcare applications, monitoring of neuromuscular impairments, and activity recognition. There is a growing interest for developing technologies and methods for enabling human motion analysis, ranging from specialized in situ systems to low-cost wearable systems.

Activity recognition, a key component in pervasive healthcare monitoring, relies on classification algorithms that require labeled data of individuals performing the activity of interest to train accurate models. Labeling data can be performed in a lab setting where an individual enacts the activity under controlled conditions. The ubiquity of mobile and wearable sensors allows the collection of large datasets from individuals performing activities in naturalistic conditions. Gathering accurate data labels for activity recognition is typically an expensive and time-consuming process. In this paper we present two novel approaches for semi-automated online data labeling performed by the individual executing the activity of interest. The approaches have been designed to address two of the limitations of self-annotation: (i) The burden on the user performing and annotating the activity, and (ii) the lack of accuracy due to the user labeling the data minutes or hours after the completion of an activity. The first approach is based on the recognition of subtle finger gestures performed in response to a data-labeling query. The second approach focuses on labeling activities that have an auditory manifestation and uses a classifier to have an initial estimation of the activity, and a conversational agent to ask the participant for clarification or for additional data. Both approaches are described, evaluated in controlled experiments to assess their feasibility and their advantages and limitations are discussed. Results show that while both studies have limitations, they achieve 80% to 90% precision.

The purpose of this study is to develop a system capable of performing calculation of temporal gait parameters using two low-cost wireless accelerometers and artificial intelligence-based techniques as part of a larger research project for conducting human gait analysis. Ten healthy subjects of different ages participated in this study and performed controlled walking tests. Two wireless accelerometers were placed on their ankles. Raw acceleration signals were processed in order to obtain gait patterns from characteristic peaks related to steps. A Bayesian model was implemented to classify the characteristic peaks into steps or nonsteps. The acceleration signals were segmented based on gait events, such as heel strike and toe-off, of actual steps. Temporal gait parameters, such as cadence, ambulation time, step time, gait cycle time, stance and swing phase time, simple and double support time, were estimated from segmented acceleration signals. Gait data-sets were divided into two groups of ages to test Bayesian models in order to classify the characteristic peaks. The mean error obtained from calculating the temporal gait parameters was 4.6%. Bayesian models are useful techniques that can be applied to classification of gait data of subjects at different ages with promising results

Activity recognition is one of the most active areas of research in ubiquitous computing. In particular, gait activity recognition is useful to identify various risk factors in people’s health that are directly related to their physical activity. One of the issues in activity recognition, and gait in particular, is that often datasets are unbalanced (i.e., the distribution of classes is not uniform), and due to this disparity, the models tend to categorize into the class with more instances. In the present study, two methods for classifying gait activities using accelerometer and gyroscope data from a large-scale public dataset were evaluated and compared. The gait activities in this dataset are: (i) going down an incline, (ii) going up an incline, (iii) walking on level ground, (iv) going down stairs, and (v) going up stairs. The proposed methods are based on conventional (shallow) and deep learning techniques. In addition, data were evaluated from three data treatments: original unbalanced data, sampled data, and augmented data. The latter was based on the generation of synthetic data according to segmented gait data. The best results were obtained with classifiers built with augmented data, with F-measure results of 0.812 ($\sigma$ = 0.078) for the shallow learning approach, and of 0.927 ($\sigma$ = 0.033) for the deep learning approach. In addition, the data augmentation strategy proposed to deal with the unbalanced problem resulted in increased classification performance using both techniques.