香京julia种子在线播放

Falls may result in key physiological variations of the human body, which provide a criterion to detect a fall. By measuring various human body related attributes using accelerometers, gyroscopes, glucometers, pressure sensors, ECG (Electrocardiography), EEG (Electroencephalography), or EOG (Electromyography), one can detect anomalies within subjects. Due to the advantages of mobility, portability, low cost, and availability, wearable devices are regarded as one of the key types of sensors for fall detection and have been widely studied. Numerous studies have been conducted to investigate wearable devices, which are regarded as a promising direction to study fall detection and prediction.

Based on our search criteria and filtering strategy (Tables 1, 2), 28 studies, including eight papers focusing on public data sets, focusing on fall detection by individual wearable devices are selected and described to illustrate trends and challenges of fall detection during the past 6 years. Some conclusions can be drawn based on the literature during the past 6 years in comparison to the studies before 2014. From Table 3, we note that studies applying accelerometers account for a large percentage of research in this field. To the best of our knowledge, only Xi et al. (2017) deployed electromyography to detect falls, and 19 out of 20 papers applied an accelerometer to detect falls. Although the equipment used, such as Shimmer nodes, smartphones, and smart watches, often contain other sensors like gyroscopes and magnetometers, these sensors were not used to detect falls. Bourke et al. (2007) also found that accelerometers are regarded as the most popular sensors for fall detection mainly due to its affordable cost, easy installation and relatively good performance.

Although smartphones have gained attention for studying falls, the underlying sensors of systems using them are still accelerometers and gyroscopes (Shi et al., 2016; Islam et al., 2017; Medrano et al., 2017; Chen et al., 2018). Users are more likely to carry smartphones all day rather than extra wearable devices, so smartphones are useful for eventual real-world deployments (Zhang et al., 2006; Dai et al., 2010).

Vision-based detection is another prominent method. Extensive effort in this direction has been demonstrated, and some of which (Akagündüz et al., 2017; Ko et al., 2018; Shojaei-Hashemi et al., 2018) show promising performance. Although most cameras are not as portable as wearable devices, they offer other advantages which deem them as decent options depending upon the scenario. Most static RGB cameras are not intrusive and wired hence there is no need to worry about battery limitations. Work on demonstrating viability of vision-based approaches have been demonstrated which makes use of infrared cameras (Mastorakis and Makris, 2014), RGB cameras (Charfi et al., 2012), and RGB-D depth cameras (Cai et al., 2017). One main challenge of vision-based detection is the potential violation of privacy due to the levels of detail that cameras can capture, such as personal information, appearance, and visuals of the living environment.

Further to the information that we report in Table 4, we note that RGB, depth, and infrared cameras are the three main visual sensors used. Moreover, it can be noted that the RGB-D camera (Kinect) is among the most popular vision-based sensor, as 12 out of 22 studies applied it in their work. Nine out of the other 10 studies used RGB cameras including cameras built into smartphones, web cameras, and monocular cameras, while the remaining study used an infrared camera within Kinect, to conduct their experiments.

Static RGB cameras are the most widely used sensors within the vision-based fall detection research conducted before 2004, although the accuracies of RGB camera-based detection systems vary drastically due to environmental conditions, such as illumination changes—which often results in limitations during the night. Besides, RGB cameras are inherently likely to have a higher false alarm rate because some deliberate actions like lying on the floor, sleeping or sitting down abruptly are not easily distinguished by frames captured by RGB cameras. With the launch of the Microsoft Kinect, which consists of an RGB camera, a depth sensor, and a multi-array microphone, it stimulated a trend in 3D data collection and analysis, causing a shift from RGB to RGB-D cameras. Kinect depth cameras took the place of the traditional RGB cameras and became the second popular sensors in the field of fall detection after 2014 (Xu et al., 2018).

In the last years, we are seeing an increased interest in the use of wearable cameras for the detection of falls. For instance, Ozcan and Velipasalar (2016) tried to exploit the cameras on smartphones. Smartphones were attached to the waists of subjects and their inbuilt cameras were used to record visual data. Ozcan et al. (2017) investigated how web cameras (e.g., Microsoft LifeCam) attached to the waists of subjects can contribute to fall detection. Although both approaches are not yet practical to be deployed in real applications, they show a new direction, which combines the advantages of wearable and visual sensors.

Table 4 reports the work conducted for individual vision-based sensors. The majority of research still makes use of simulated data. Only two studies use real world data; the one by Boulard et al. (2014) has actual fall data and the other by Stone and Skubic (2015) has mixed data, including 9 genuine falls and 445 simulated falls by trained stunt actors. In contrast to the real data sets from the work of Klenk et al. (2016) collected by wearable devices, there are few purely genuine data sets collected in real life scenarios using individual visual sensors.

The ambient sensor provides another non-intrusive means of fall detection. Sensors like active infrared, RFID, pressure, smart tiles, magnetic switches, Doppler Radar, ultrasonic, and microphone are used to detect the environmental changes due to falling as shown in Table 5. It provides an innovative direction in this field, which is passive and pervasive detection. Ultra-sonic sensor network systems are one of the earliest solutions in fall detection systems. Hori et al. (2004) argues that one can detect falls by putting a series of spatially distributed sensors in the space where elderly persons live. In Wang et al. (2017a,b), a new fall detection approach which uses ambient sensors is proposed. It relies on Wi-Fi which, due to its non-invasive and ubiquitous characteristics, is gaining more and more popularity. However, the studies by Wang et al. (2017a,b) are limited in terms of multi-person detection due to their classifiers not being robust enough to distinguish new subjects and environments. In order to tackle this issue, other studies have developed more sophisticated methods. These include the Aryokee (Tian et al., 2018) and FallDeFi (Palipana et al., 2018) systems. The Aryokee system is ubiquitous, passive and uses RF-sensing methods. Over 140 people were engaged to perform 40 kinds of activities in different environments for the collection of data and a convolutional neural network was utilized to classify falls. Palipana et al. (2018) developed a fall detection technique named FallDeFi, which is based on WiFi signals as the enabling sensing technology. They provided a system applying time-frequency of WiFi Channel State Information (CSI) and achieved above 93% average accuracy.

RF-sensing technologies have also been widely applied to other recognition activities beyond fall detection (Zhao et al., 2018; Zhang et al., 2019) and even for subtle movements. Zhao et al. (2018) studied human pose estimation with multiple persons. Their experiment showed that RF-pose has better performance under occlusion. This improvement is attributable to the ability of their method to estimate the pose of the subject through a wall, something that visual sensors fail to do. Further research on RF-sensing was conducted by Niu et al. (2018) with applications to finger gesture recognition, human respiration and chins movement. Their research can be potentially used for applications of autonomous health monitoring and home appliances control. Furthermore, Zhang et al. (2019) used an RF-sensing approach in the proposed system WiDIGR for gait recognition. Guo et al. (2019) claimed that RF-sensing is drawing more attention which can be attributed to being device-free for users, and in contrast to RGB cameras it can work under low light conditions and occlusions.

For most research groups there is not enough time and funding to collect data continuously within several years to study fall detection. Due to the rarity of genuine data in fall detection and prediction, Li et al. (2013) have started to hire stunt actors to simulate different kinds of fall. There are also many data sets of falls which are simulated by young healthy students as indicated in the studies by Bourke et al. (2007) and Ma et al. (2014). For obvious reasons elderly subjects cannot be engaged to perform the motion of falls for data collection. For most of the existing data sets, falls are simulated by young volunteers who perform soft falls under the protection of soft mats on the ground. Elderly subjects, however, often have totally different behavior due to less control over the speed of the body. One potential solution could include simulated data sets created using physics engines, such as OpenSim. Previous research (Mastorakis et al., 2007, 2018) have shown that simulated data from OpenSim contributed to an increase in performance to the resulting models. Another solution includes online learning algorithms which adapt to subjects who were not represented in the training data. For instance, Deng et al. (2014) applied the Transfer learning reduced Kernel Extreme Learning Machine (RKELM) approach and showed how they can adapt a trained classifier—based on data sets collected by young volunteers—to the elderly. The algorithm consists of two parts, namely offline classification modeling and online updating modeling, which is used to adapt to new subjects. After the model is trained by labeled training data offline, unlabeled test samples are fed into the pre-trained RKELM classifier and obtain a confidence score. The samples that obtain a confidence score above a certain threshold are used to update the model. In this way, the model is able to adapt to new subjects gradually when new samples are received from new subjects. Namba and Yamada (2018a,b) demonstrated how deep reinforcement learning can be applied to assisting mobile robots, in order to adapt to conditions that were not present in the training set.

Threshold-based and data-driven algorithms (including machine learning and deep learning) are the two main approaches that have been used for fall detection. Threshold-based approaches are usually used for data coming from individual sensors, such as accelerometers, gyroscopes, and electromyography. Their decisions are made by comparing measured values from concerned sensors to empirically established threshold values. Data driven approaches are more applicable for sensor fusion as they can learn non-trivial non-linear relationships from the data of all involved sensors. In terms of the algorithms used to analyze data collected using wearable devices, Figure 5 demonstrates that there is a significant shift to machine learning based approaches, in comparison to the work conducted between 1998 and 2012. From papers presented between 1998 and 2012, threshold-based approaches account for 71%, while only 4% applied machine learning based methods (Schwickert et al., 2013). We believe that this shift is due to two main reasons. First, the rapid development of affordable sensors and the rise of the Internet-of-Things made it possible to more easily deploy multiple sensors in different applications. As mentioned above the non-linear fusion of multiple sensors can be modeled very well by machine learning approaches. Second, with the breakthrough of deep learning, threshold-based approaches have become even less preferable. Moreover, different types of machine learning approaches have been explored, namely, Bayesian networks, rule-based systems, nearest neighbor-based techniques, and neural networks. These data-driven approaches (Gharghan et al., 2018) show better accuracy and they are more robust in comparison to threshold-based methods. Notable is the fact that data-driven approaches are more resource hungry than threshold-based methods. With the ever advancement of technology, however, this is not a major concern and we foresee that more effort will be invested in this direction.

Traditional machine learning approaches determine mapping functions between extracted handcrafted features from raw training data and the respective output labels (e.g., no fall or fall, to keep it simple). The extraction of handcrafted features requires domain expertise and are, therefore, limited to the knowledge of the domain experts. Though such a limitation is imposed, literature shows that traditional machine learning, based on support vector machines, hidden Markov models, and decision trees are still very active in the field of fall detection that uses individual wearable non-visual or ambient sensors (e.g., accelerometer) (Wang et al., 2017a,b; Chen et al., 2018; Saleh and Jeannès, 2019; Wu et al., 2019). For visual sensors the trend has been moving toward deep learning for convolutional neural networks (CNN) (Adhikari et al., 2017; Kong et al., 2019; Han et al., 2020), or LSTM (Shojaei-Hashemi et al., 2018). Deep learning is a sophisticated learning framework that besides the mapping function (as mentioned above and used in traditional machine learning), it also learns the features (in a hierarchy fashion) that characterize the concerned classes (e.g., falls and no falls). This approach has been inspired by the visual system of the mammalian brain (LeCun et al., 2015). In computer vision applications, which take as input images or videos, deep learning has been established as state-of-the-art. In this regard, similar to other computer vision applications, fall detection approaches that rely on vision data have been shifting from traditional machine learning to deep learning in recent years.

Real-time is a key feature for fall detection systems, especially for commercial products. Considering that certain falls can be fatal or detrimental to the health, it is crucial that the deployed fall detection systems have high computational efficiency, preferably operating in (near) real-time. Below, we comment how the methods proposed in the reviewed literature fit within this aspect.

The percentage of studies applying real-time detection by static visual sensors are lower than that of wearable devices. For the studies using wearable devices, Table 3 illustrates that six out of 20 studies that we reviewed can detect falls and send alarms. There are, however, few studies which demonstrate the ability to process data and send alerts in real-time for work conducted using individual visual sensors. Based on Table 4, one can note that although 40.9% (nine out of 22) of the studies claim that their systems can be used in real-time only one study showed that an alarm can actually be sent in real-time. The following are a couple of reasons why a higher percentage of vision-based systems can not be used in real time. Firstly, visual data is much larger and, therefore, its processing is more time consuming than that of one dimensional signals coming from non-vision-based wearable devices. Secondly, most of the work using vision sensors conducted their experiments with off-line methods, and modules like data transmission were not involved.

In terms of sensor fusion, there are two categories, typically referred to as homogeneous and heterogeneous which take input from three types of sensors, namely wearable, visual, ambient sensors, as shown in Figure 6. Sensor fusion involves using multiple and different signals coming from various devices, which may for instance include, accelerometer, gyroscope, magnetometer, and visual sensors, among others. This is all done to complement the strengths of all devices for the design and development of more robust algorithms that can be used to monitor the health of subjects and detect falls (Spasova et al., 2016; Ma et al., 2019).

For the visual detection based approaches, the fusion of signals coming from RGB (Charfi et al., 2012), and RGB-D depth cameras along with camera arrays have been studied (Zhang et al., 2014). They showed that such fusion provides more viewpoints of detected locations, and improves the stability and robustness by decreasing false alarms due to occluded falls (Auvinet et al., 2011).

Li et al. (2018) combined accelerometer data from smartphones and Kinect depth data as well as smartphone camera signals. Liu et al. (2014) and Yazar et al. (2014) fused data from infrared sensors with ambient sensors, and data from doppler and vibration sensors separately. Among them, accelerometers and depth cameras (Kinect) are most frequently studied due to their low costs and effectiveness.

Most of the existing IoT platforms, such as Microsoft Azure IoT, IBM Watson IoT Platform, and Google Cloud Platform, have not been used in the deployment of fall detection approaches by sensor fusion. In general, research studies on fall detection using sensor fusion are carried out by offline methods and decision fusion approaches. Therefore, in such studies, there is no need for data transmission and storage modules. From Tables 6, 7, one can also observe that most of the time researchers applied their own workstations or personal computers as their platforms, as there was no need for the integration of sensors and real-time analysis in terms of fall detection in off-line mode.

Some works, such as those in Kwolek and Kepski (2014), Kepski and Kwolek (2014), and Kwolek and Kepski (2016), applied low-power single-board computer development platforms running in Linux, namely PandaBoard, PandaBoard ES, and A13-OlinuXino. A13-OlinuXino is an ARM-based single-board computer development platform, which runs Debian Linux distribution. PandaBoard ES, which is the updated version of PandaBoard, is a single-board computer development platform running at Linux. The PandaBoard ES can run different kinds of Linux-based operating systems, including Android and Ubuntu. It consists of 1 GB of DDR2 SDRAM, dual USB 2.0 ports as well as wired 10/100 Ethernet along with wireless Ethernet and Bluetooth connectivity. Linux is well-known for real-time embedded platforms since it provides various flexible inter-process communication methods, which is quite suitable for fall detection using sensor fusion.

In the research by Kwolek and Kepski (2014, 2016), wearable devices and Kinect were connected to the Pandaboard through Bluetooth and cable, separately. Firstly, data was collected by accelerometers and Kinect sensors, individually, which was then transmitted and stored in a memory card. The procedure of data transmission is asynchronous since there are different sampling rates for accelerometers and Kinect. Finally, all data was grouped together and processed by classification models that detected falls. The authors reported high accuracy rates but could not compare with other approaches since there is no benchmark data set.

Spasova et al. (2016) applied the A13-OlinuXino board as their platform. A standard web camera was connected to it via USB and an infrared camera was connected to the development board via I2C (Inter-Integrated Circuit). Their experiment achieved excellent performance with over 97% sensitivity and specificity. They claim that their system can be applied in real-time with hardware of low-cost and open source software platform.

Despite the available platforms mentioned above, the majority of fall detection studies trained their models in an offline mode with a single sensor on personal computers. The studies in Kwolek and Kepski (2014), Kepski and Kwolek (2014), Kwolek and Kepski (2016), and Spasova et al. (2016) utilized single-board computer platforms in their experiments to demonstrate the efficacy of their approaches. The crucial aspects of scalability and efficiency were not addressed and hence it is difficult to speculate the appropriateness of their methods in real-world applications. We believe that the future trend is to apply an interdisciplinary approach that deploys the data analysis modules on mature cloud platforms, which can provide a stable and robust environment while meeting the exploding demands of commercial applications.

Although some groups devoted their efforts to acquire data of genuine falls, most researchers used data that contained simulated falls. We know that monitoring the lives of elderly people and waiting to capture real falls is very sensitive and time consuming. Having said that though, with regards to sensor fusion by wearable devices, there have been some attempts which have tried to build data sets of genuine data in real life. FARSEEING (Fall Repository for the design of Smart and self-adaptive Environments prolonging Independent living) is one such data set (Klenk et al., 2016). It is actually the largest data set of genuine falls in real life, and is open to public research upon request on their website. From 2012 to 2015, more than 2,000 volunteers have been involved, and more than 300 real falls have been collected under the collaboration of six institutions³.

As for the fusion by visual sensors and the combination of other non-wearable sensors, it becomes quite hard to acquire genuine data in real life. There was one group which tried to collect real data by visual sensors, but only nine real falls by elderly (Demiris et al., 2008) were captured during several years. The availability of only nine falls is too limited to train a meaningful model. As an alternative, Stone and Skubic (2015) hired trained stunt actors to simulate different kinds of falls and made a benchmark data set with 454 falls including 9 real falls by elderly.

Speaking of the fusion of different sensors, there are several criteria to group them. Yang and Yang (2006) and Tsinganos and Skodras (2018) grouped them into three categories, namely direct data fusion, feature fusion, and decision fusion. We divide sensor fusion techniques into four groups as shown in Figure 7, which we refer to as fusion with partial sensors, direct data fusion, feature fusion, and decision fusion.

For the partial fusion, although multiple sensors are deployed, only one sensor is used to take the final decision, such as the work by Ma et al. (2019). They used an RGB and a thermal camera to conduct their experiments, with the thermal camera being used only for the localization of faces. Falls were eventually detected only based on the data collected from the regular RGB cameras. A similar approach was applied by Spasova et al. (2016), where an infrared camera was deployed to confirm the presence of the subject and the data produced by the RGB camera was used to detect falls. There are also other works that used wearable devices that deployed the sensors at different stages. For instance, in (Kepski and Kwolek, 2014; Kwolek and Kepski, 2014) a fall detection system was built by utilizing a tri-axial accelerometer and an RGB-D camera. The accelerometer was deployed to detect the motion of the subject. If the measured signal exceeded a given threshold then the Kinect was activated to capture the ongoing event.

The second approach of sensor fusion is known as feature fusion. In such an approach, feature extraction takes places on signals that come from different sensors. Then all features are merged into long feature vectors and used to train classification models. Most of the studies that we reviewed applied feature fusion for wearable-based fall detection systems. Many commercial products of wearable devices, sensors like accelerometers, gyroscope, magnetometer are built in one device. Data from these sensors is homogeneous synchronous with the same frequency and transmitted with built-in wireless modules. Having signals producing data with the synchronized frequency simplifies the fusion of data. Statistical features, such as mean, maximum, standard deviation, correlation, spectral entropy, spectral, sum vector magnitude, the angle between y-axis and vertical direction, and differential sum vector magnitude centroid can be determined from the signals of accelerometers, magnetometers, and gyroscopes, and used as features to train a classification model that can detect different types of falls (Yang et al., 2016; de Quadros et al., 2018; Gia et al., 2018).

Decision fusion is the third approach, where a chain of classifiers is used to come to a decision. A typical arrangement is to have a classification model that takes input from one type of sensor, another model that takes input from another sensor, and in turn the outputs of these two models are used as input to a third classification model that takes the final decision. Li et al. (2018) explored this approach with accelerometers embedded in smart phones and Kinect sensors. Ozcan and Velipasalar (2016) deployed an accelerometer and an RGB camera for the detection of falls. Different sensors, such as accelerometer, RGB and RGB-D cameras were deployed in these studies. Decisions are made separately based on the individual sensors, and then the final decision is achieved by combining the individual sensors.

The final approach is data fusion. This is achieved by first fusing the data from different sensors and perform feature extraction from the fused data. This is in contrast to feature fusion where data from these sensors is homogeneous synchronous with the same frequency. Data fusion can be applied to different sensors with different sampling frequency and data characteristics. Data from various sensors can be synchronized and combined directly for some sensors of different types. Because of the difference in sampling rate between the Kinect camera and wearable sensors, it is challenging to conduct feature fusion directly. In order to mitigate this difficulty, the transmission and exposure times of the Kinect camera are adapted to synchronize the RGB-D data with that of wearable sensors by an ad-hoc acquisition software, as was done by Gasparrini et al. (2015).

Ozcan and Velipasalar (2016) used both partial and feature fusion. They divided the procedure in two stages. In the first stage, only the accelerometer was utilized to indicate a potential fall, then the Kinect camera activates after the accelerometer flagged a potential fall. Features from both the Kinect camera and accelerometer were then extracted to classify activities of fall or non-fall in the second stage.

In terms of fall detection techniques based on wearable sensor fusion, the explored methods include threshold-based, traditional machine learning, and deep learning. The latter two are the most popular due to their robustness. The research by Chelli and Pätzold (2019) applied both traditional machine learning [kNN, QSVM, Ensemble Bagged Tree (EBT)] and deep learning. Their experiments were divided into two parts, namely activity recognition and fall detection. For the former, their experiments showed that traditional machine learning and deep learning outperformed other approaches, which showed 94.1 and 93.2% accuracy, respectively. Queralta et al. (2019) applied a long short-term memory (LSTM) approach, where wearable nodes including accelerometer, gyroscope, and magnetometer were embedded in a low power wide area network, with combined edge and fog computing. The LSTM algorithm is a type of recurrent neural network aimed at solving long sequence learning tasks. Their system achieved an average recall of 95% while providing a real-time solution of fall detection running on cloud platforms. Another example is the work by Nukala et al. (2014) who fused the measurements of accelerometers and gyroscopes and applied an Artificial Neural Network (ANN) for the modeling of fall detection.

As for visual sensor based fusion techniques, the limited studies that were included in our survey applied either traditional machine learning or deep learning (Espinosa et al., 2019; Ma et al., 2019) approaches. Fusion of multiple visual sensors from a public data set was presented by Espinosa et al. (2019), where a 2D CNN was trained to classify falls during daily life activities.

Another approach is reinforcement learning (RL), which is a growing branch in machine learning, and is gaining popularity in the fall detection field as well. Deep reinforcement learning (DRL) combines the advantages of deep learning and reinforcement learning, and has already shown its benefits in fall prevention (Namba and Yamada, 2018a,b; Yang, 2018) and fall detection (Yang, 2018). Namba and Yamada (2018a) proposed a fall risk prevention approach by assisting robots for the elderly living independently. Images and movies with the location information of accidents were collected. Most conventional machine learning and deep learning methods are, however, challenged when the operational environment changes. This is due to their data-driven nature that allows them to learn how to become robust mostly in the same environments where they were trained.

Typical data storage devices include SD cards, local storage on the integration device, or remote storage on the cloud. For example, some studies used the camera and accelerometer in smartphones, and stored the data on the local storage of the smarphones (Ozcan and Velipasalar, 2016; Shi et al., 2016; Medrano et al., 2017). Other studies applied off-line methods and stored data in their own computer, and could be processed at a later stage. Alamri et al. (2013) argue that sensor-cloud will become the future trend because cloud platforms can be more open and more flexible than local platforms, which have limited local storage and processing power.

There seems to be a general consensus that sensor fusion provides a more robust approach for the detection of elderly falls. The use of various sensors may complement each other in different situations. Thus, instead of relying on only one sensor, which may be unreliable if the conditions are not suitable for that sensor, the idea is to rely on different types of sensor that together can capture reliable data in various conditions. This results in a more robust system that can keep false alarms to a minimum while achieving high precision.

Conventional machine learning approaches have been widely applied in fall detection and activity recognition, and results outperform those of threshold-based methods in studies that use wearable sensors. Deep learning is a subset of machine learning, which is concerned with artificial neural networks inspired by the mammalian brain. Approaches of deep learning are gaining popularity especially for visual sensors and sensor fusion and are becoming the state-of-the-art for fall detection and other activity recognition. Deep reinforcement learning is another promising research direction for fall detection. Reinforcement learning is inspired by the psychological neuro-scientific understandings of humans which can adapt and optimize decisions in a changing environment. Deep reinforcement learning combines advantages of deep learning, and reinforcement learning which can provide alternatives for detection that can adapt to the changing condition without sacrificing accuracy and robustness.

5G is a softwarized and virtualized wireless network, which includes both a physical network and software virtual network functions. In comparison to 4G networks, 5th generation mobile introduces the ability of data transmission with high speed and low latency, which could contribute to the development of fall detection by IoT systems. Firstly, 5G is envisioned to become an important and universal communication protocol for IoT. Secondly, 5G cellular can be used for passive sensing approaches. Different from other kinds of RF-sensing approaches (e.g., WiFi or radar) which are aimed for short-distance indoor fall detection, the 5G wireless network can be applied to both indoor and outdoor scenarios as a pervasive sensing method. This type of network has already been successfully investigated by Gholampooryazdi et al. (2017) for the detection of crowd-size, presence detection, and walking speed, and their experiments showed accuracy of 80.9, 92.8, and 95%, respectively. Thirdly, we expect that 5G as a network is going to become a highly efficient and accurate platform to achieve better performance of anomaly detection. Smart networks or systems powered by 5G IoT and deep learning can be applied not only in fall detection systems, but also in other pervasive sensing and smart monitoring systems which assist elderly groups to live independently with high-quality life.

El-Bendary et al. (2013) and Namba and Yamada (2018b) have proposed to include historical medical and behavioral data of individuals along with sensor data. This allowed the enrichment of the data and consequently to make better informed decisions. This innovative perspective allows a more personalized approach as it uses the health profile of the concerned individual and it has the potential to become a trend also in this field. Another trend could be the way data sets are created to evaluate systems for fall detection. Mastorakis et al. (2007, 2018) applied the skeletal model simulated in Opensim, which is an open-source software developed by Stanford University. It can simulate different kinds of pre-defined skeletal models. They acquired 132 videos of different types of falls, and trained their own algorithms based on those models. The high results that they report indicate that the simulated falls by OpenSim are very realistic and, therefore, effective for training a fall detection model. Physics engines, like Opensim, can simulate customized data based on the height and age of different subjects and it offers the possibility of new directions to detect falls. Another solution, which can potentially address the scarcity of data, is to develop algorithms that can be adapted to subjects that were not part of the original training set (Deng et al., 2014; Namba and Yamada, 2018a,b) as we described in section 4.1.4.

As to architecture is concerned, Fog computing offers the possibility to distribute different levels of processing across the involved edge devices in a decentralized way. Smart devices that can carry out some processing and that can communicate directly with each other are more attractive for (near) real-time processing as opposed to systems based on cloud computing (Queralta et al., 2019). An example of such smart devices include the Intel® RealSense™ depth camera, which includes a 28 nanometer (nm) processor to compute real-time depth images.