Human monitoring plays an important role in smart homes, such as in assisted living scenarios or emergency detection. A fundamental problem in human monitoring is how to localize humans in indoor environments. In [
1], a localization method was proposed for tracking humans’ position in indoor environments based on Passive Infrared (PIR) sensors. First, the grid-based accessibility map, which reflects human visiting preferences and the physical layout of the area, is built. Then, PIR sensors, deployed according to the grid-based accessibility map, provide an external rough position of the human. Finally, an A-star algorithm fuses the PIR sensor data and grid-based accessible map information to estimate the trajectory and reduce the errors. Experiments have been performed in a mock apartment testbed, and the method could prove especially useful in service robotics applications. The position of people within a smart home can then be used as the base for more complex services. For instance, the work in [
2] proposed a smart terminal for remote “one-click” control of devices, driven by the position of the user estimated via fingerprint matching based on received signal strength analysis, together with pedestrian dead reckoning. The solution has been tested in a laboratory scenario. Human monitoring, in its deeper sense, aims to recognize important events, especially those that may hinder people’s safety. Fall detection is one of the most investigated events in assistive solutions. The authors in [
3] proposed an accurate fall detection method investigating the depth frames of the human body using a single device in a top-view configuration, with the subjects located under the device inside a room. Features extracted from depth frames train a classifier based on a binary support vector machine learning algorithm. A fall is identified when the distance between the Kinect and the centroid associated with the person’s head becomes comparable with the floor distance. Comprehensive monitoring, however, requires using several sensor (a so-called sensor network) and inferring useful information from a noticeable amount of data. In order to exploit such information, however, powerful analytics is needed to convert raw sensor output into meaningful and accessible knowledge. In [
4], a complete monitoring architecture was presented, which relies on a cloud-based approach. Behavioral explanatory models are introduced, with the purpose of quantitatively monitoring a given quantity of interest (e.g., toilet visits across time). Then, a tool that models the probability of the activation of given sensors throughout the day is presented: this enables a statistical hypothesis testing frame for detecting changes between two different time periods. Finally, a methodology to extract complex user patterns was presented, based on machine-learning techniques to infer meaningful information from sensor data, thus dealing with the inherent variability of human behaviors. The system has been deployed at several pilot sites and tested on real data. The authors of [
5], instead, focused more on the smart sensing architecture, presenting an integrated sensor network to monitor the user and the environment in order to derive information about the user’s behavior and her/his health status. The proposed platform includes biomedical, wearable, and unobtrusive sensors for monitoring user’s physiological parameters and home automation sensors to obtain information about her/his environment. The sensor network stores the heterogeneous data both locally and remotely in the cloud, where machine learning algorithms and data mining strategies are used for user behavior identification, classification of user health conditions, classification of the smart home profile, and data analytics to implement services for the community. The proposed solution has been experimentally tested in a pilot study based on the development of both sensors and services for elderly users at home. Whenever monitoring is performed via cameras, the problem of “privacy” arises. The authors of [
6] investigated the detection of embarrassing situations for social robots in smart homes using convolutional neural networks. The paper aimed to protect the sensitive information at the beginning of video data collection, thanks to a developed mechanism, which permits a social robot to detect embarrassing situations and convert privacy information to non-sensitiveinformation. This is done via an improved neural network structure and feature extraction algorithms based on You Only Look Once (YOLO). The algorithm has been trained with a dataset of six classes of situations in the smart home: taking a shower, sleeping (naked or half-naked), using the toilet, dressing (naked or half-naked), humans are in the smart home, but no privacy context is involved, and no person in the smart home. Tests were performed on different sets of images.