**4. Seniors' safety monitoring: a case study of SHAH**

The entire world is witnessing a fast-growing aging population body, and it is widely observed that there are more and more seniors living alone, either in their homes or in individual rooms of nursing houses. An effective healthcare service system is required to maintain 24/7 real-time monitoring and timely dangerous action recognition. The compelling need for both regular medical consultation and timely emergency assistance inspires us to adopt this topic for our case study.

#### **4.1 Sensor types**

There are two types of sensors that are normally used in action recognition: wearable sensors and remote sensors. Wearable sensors are compact and lightweight, making them convenient for real-world settings. They can provide detailed motion data, including acceleration, orientation, and angular velocity information. This type of data helps to recognize specific actions, such as walking, standing, and sitting. However, wearable sensors can be limited in terms of the scope of their coverage. They may not be able to capture specific actions or movements that occur outside of the sensor's range or when the sensor is not worn.

Remote sensors, such as cameras, can capture visual data that complements the motion data provided by wearables. This type of data gives a more comprehensive view of the environment and enables the recognition of complex actions that may not be easily detectable from motion data alone. For example, cameras can capture facial expressions, gestures, and interactions with objects in the environment. However, remote sensors can be limited by lighting conditions, occlusion, and the need for a clear line of sight.

By combining the advantages of both wearable and remote sensors, it is possible to develop a more comprehensive action recognition system that can perform real-time and accurate recognition of a wide range of actions in different settings. In addition, data fusion is a technique used to integrate data from multiple sources to improve the accuracy and reliability of action recognition systems, which can enhance the effectiveness of healthcare and other applications.

#### **4.2 Information fusion**

Information fusion is an approach to integrating data from multiple sources to improve the accuracy and comprehensiveness of the information obtained. It is typically divided into three levels: data level, feature level, and decision level [13].

#### *Smart Healthcare at Home in the Era of IoMT DOI: http://dx.doi.org/10.5772/intechopen.113208*

At the data level, raw data from sensors, such as wearable devices and cameras, are combined to represent the phenomenon being monitored thoroughly. This approach can be computationally intensive and requires careful calibration and synchronization of the sensors, but it can provide a more comprehensive understanding of the phenomenon.

At the feature level, features extracted from the raw data are combined to obtain a more informative and complementary representation of the phenomenon. This approach can be more efficient than data-level fusion but requires careful selection and processing of the features to ensure they are informative and complementary.

At the decision level, the decisions or outputs of different classifiers are combined to make a final decision about the phenomenon being monitored. This approach can be used when the various sources of information provide redundant information but are not necessarily complementary. It can also be used to weigh the different sources of data according to their reliability or importance.

In the IoT context, data-level fusion has been preferred due to its ability to integrate data from different sources and provide a more accurate representation of the physical environment. After the data from various sensors are fused at the data level, the resulting data is typically more compact and comprehensive than the raw data from each sensor, as shown in **Figure 3**. Consequently, the amount of data to be transmitted is less than the sample data combination, which is beneficial in conserving network bandwidth and reducing power consumption. In IoT-based action recognition, the fused data can be uploaded to the cloud or processed locally using fewer resources, such as computing power, memory, and energy. By minimizing the resources needed for uploading and processing, the overall system can operate more efficiently and cost-effectively while still achieving high accuracy and real-time performance.

#### **4.3 Data processing methodology**

Singular Spectral Analysis (SSA) is a powerful signal-processing technique that has been applied to various domains, including action recognition. In the context of IoTbased action recognition, SSA can be used to implement the skeleton data by combining wearable data [14].

In SSA, time-series data is first embedded into a trajectory matrix, where each row represents a trajectory or a subsequent trajectory of the original data. The trajectory matrix is then decomposed using singular value decomposition (SVD) to separate the data into singular values and corresponding singular vectors. These singular vectors represent the fundamental building blocks of time series. Relevant patterns and features can be extracted by selecting specific singular vectors and their associated singular values.

**Figure 3.** *Data-level information fusion.*

The fundamental components obtained by SSA can be interpreted as different patterns of variation in the time series. They capture underlying trends and recurring patterns in the data. These components can be further analyzed and combined to reconstruct the original time series or for various applications such as noise reduction, feature extraction, or forecasting. SSA provides a flexible and practical approach to analyzing time series data and has applications in multiple fields such as finance, climate science, and signal processing.

**Figure 4** shows the first ten components sorted by singular value, which is the analysis result of the example of accelerator data in the X-axis when falling. The components are clear enough to represent the trade of initial sensors, which can be used to implement the skeleton data, as it is shown in **Figure 5**.

#### **4.4 Experimental results**

Two databases are used in the experimental study. To represent the skeleton data, the NTURGB+D database is chosen. The NTURGB+D (NTU RGB + D) database is a comprehensive benchmark dataset widely used for human action recognition and pose estimation. It comprises a total of 56,880 action samples, recorded from 40 subjects performing 80 distinct actions. Each action has 20 instances, resulting in a diverse set of data for analysis. The database includes RGB videos, depth maps, and skeleton data, providing rich multi-modal information for studying human activities.

**Figure 4.** *Accelerator data in X-axis for falling and sequence of the first ten components sorted by singular value.*

**Figure 5.** *SSA-based skeleton data implementation.*

#### *Smart Healthcare at Home in the Era of IoMT DOI: http://dx.doi.org/10.5772/intechopen.113208*

To represent the initial data, the SCUT-NAA dataset is used. It is a 3D accelerationbased activity dataset that consists of 1278 samples collected from 44 individuals (34 males and 10 females). The data was gathered in naturalistic settings using a single tri-axial accelerometer placed in three different locations: waist belt, pants pocket, and cloth pocket. Each participant was asked to perform ten activities, providing a diverse range of motion data for analysis. To enhance the falling data, the 2015 Fall dataset is added, where data was collected from 32 volunteers specifically focusing on falls. The dataset includes four fall postures: forward, backward, left, and right. Sensors were primarily placed on the chest and thighs to capture both acceleration and angular velocity data during falls. This dataset offers valuable insights into different fall scenarios, enabling researchers to study and develop effective fall detection and prevention algorithms.

After selecting common actions from the above databases and sorting them out, we obtained an experimental dataset with eight actions: Sitting, Walking, Step walking, Jumping, Upstairs, Downstairs, Cycling, and Falling. The total dataset number is 879.

The falling detection result achieved an impressive accuracy rate of 93.82% using the SSA implemented on skeleton data. This outcome demonstrates the effectiveness of the SSA approach in accurately identifying and detecting falls in the dataset. By analyzing the spatiotemporal patterns of skeletal movements, the SSA-based method successfully captured the distinct characteristics associated with falls, leading to highly accurate detection results. The high accuracy rate signifies the potential of SSAbased skeleton data analysis for real-time fall detection systems, which can play a crucial role in ensuring the safety and well-being of individuals, particularly seniors and those at risk of falling. The remarkable performance underscores the value of SSA in enhancing fall detection capabilities and highlights its significance in advancing research and applications in healthcare and eldercare domains.
