An Analysis of Fall Detection Technologies
Allowing the Elderly and Disabled to Remain in their Own Homes
Executive Summary
Many elderly people are at a very high risk of injury due to their increased chance of experiencing dangerous falls and not being able to access the help they need, this can limit their independence and quality of life, as well as cause stress to families and loved ones. Developing fall detection systems prevent the need for patients to call for aid themselves in the event of a fall, and therefore remove a lot of the associated anxiety and reduce the number of emergency hospital admissions. Three approaches to system development (wearable, vision-based, ambience based) are reported and SWOT analysed before the problem as a whole is also analysed, and a conclusion is presented. Wearable systems are a decent approach, but struggle to detect certain categories of falls; vision systems are obviously limited by their field of sight but can detect a range of activities; a new ambience approach shows potential but fails when multiple people are within its range. A decision to focus on the development of vision systems was made due to their ability to detect non-typical falls and the non-intrusive nature of their methods, global advances in machine learning and computer vision have the potential to contribute to the development of highly accurate prospective solutions.
Submitted for Independent Learning Module, Computer Science & AI, University of York 2018
Page Contents
Wearable Devices
Vision Systems
Ambience Approach
SWOT Analysis of Systems, generally
Helpful Definitions
False Positive(s):
When an activity similar to falling, such as sitting down quickly, is detected as a fall.
False Negative(s):
When a fall is not detected as the resulting position is similar to a normal position, such as a fall resulting in sitting on the floor.
Decision Tree:
A branching graph used to represent decision options. Weights can be assigned to outcomes to automate decision making.
1. Background
One of the major risks facing elderly, or otherwise physically vulnerable medical patients, is their increased chance of experiencing dangerous falls. In 2013, 255,000 over 65-year olds were urgently admitted to hospital as a result of falling [1]. Patients can struggle to recover themselves from falls and rely on alerting a carer. As well as being physically unable when having fallen, some patients with mental illnesses, or disabilities, may also struggle to recognise / recall that they have access to help. The speed of the response to a fall can dramatically impact the extent of the damage to the victim’s health, so locating a victim quickly is integral to minimizing the effects. In recent years the number of “emergency” hospital admissions as a result of falling has declined [see figure 1, for example]. Patients whom are found quicker can be treat less urgently and may not even require a hospital admission – it is often when patients are not found after falling that the “emergency” element of the care arises.
Many fall detection technologies have been researched and developed which aim to remove the need for the individual to call for the aid themselves and lessen the number of emergency cases. By lessening the danger that these falls pose, the technologies contribute to an elderly person's ability to remain at home, as opposed to being relocated into full-time care. By remaining at home, a patient’s perception of their quality of life and day to day satisfaction is generally better preserved. Home living situations should usually be supported by a patient's family, or similar support system, and these fall detection technologies do not serve as a substitute for care, but as a safeguarding technique. They also help to provide peace of mind for a patient and their family, reduce the strain on local emergency hospital departments and save money on treatment. With the rapid growth of the aging population, demand for these systems is increasing.
2. Introduction
There are a number of different approaches taken when tackling the issue of fall detection, ranging from the more intrusive to the somewhat unnoticeable. Each system must have a way of determining whether the change to an individual’s position has been voluntary or not, and must cover a range of instances, not just the classic transition from of being upright to horizontal but also cases when a patient has fallen on sloped environments such as stairs, or cases when the patient was already horizontal, such as falling out of bed. “A survey of Fall Detection” [2] categorized the common approaches into three distinct types of systems: wearable device, ambience device and vision based. A system is considered successful when it is proven to be highly sensitive and situationally robust. The survey concluded that although no ideal system had been developed, the existing strategies and algorithms provided a good basis for further development.
Before discussing the successes of vision solutions developed since the publication of [2] in 2013, and new advances in ambient systems, the aspects of older, wearable systems will be evaluated, as these were a very common approach which have recently been replaced by more computational complex systems in an effort to improve performance. Each approach will be analysed from a SWOT perspective and an overall SWOT analysis is included, before an overall summary is presented at the end of the report.
3.1 Wearable Devices
Wearable devices proved to be a cheap and easy to set up. In 2006, a device worn on the waist was developed which relied on the methodology of detecting an impact, and then observing the change in orientation of the wearer [3]. It aimed to improve upon previous devices which struggled to detect backwards and sideways falls [4], by incorporating the emphasis on orientation change as opposed to a multi-stage, impact-based algorithms used by other devices. Placing the device on the waist of the patient reduced the risk of false positives caused by other activities such as sitting, which when only monitoring impact can follow similar patterns to falling. By using a “differential method of orientation” [3] the algorithm used was able to detect falls where the user does not end completely horizontally, such as against a piece of furniture, an area where other orientation-based systems had struggled [5] when only taking into account the orientation of the wearer after the impact. It was also found that bending down quickly could cause false positives, as the device mistook this for a slanted fall.
An improvement on the methodology described above was documented in [6]. The system was composed of 2 tri-axial accelerometers, for measuring vibration simultaneously along three axes, placed on the patient’s chest and thigh in order to “recognize four kinds of static postures: standing, bending, sitting, and lying” The measurements taken, acceleration and angular velocity, were evaluated by the algorithm to deduce whether movement between these static postures was intentional; whether the patient has fallen or not. By this method, the number of both false positives and false negatives were lessened, when compared to earlier systems [3, 4, 5]. However, the new system was not perfect. Only a 92% success rate was achieved when the test users were to lie down quickly, delivering a false positive the other 8% of the time, and the system also struggled to register falls against a wall which resulted in sitting, delivering false negative 9% of the time. It was suggested that in order to aid the device’s decision making, a method of informing which room the patient was in would be beneficial; “if the person is in the bedroom, jumping into bed is more likely to happen” [6].
Analysis
Wearable detection devices prove to be a decent solution: typical falls are accurately detected, and they are computationally cheap which results in a real-time response. They are also relatively cheap to produce, as no complex hardware is required. By design, they all have the negative factor of being somewhat intrusive, although they are not cumbersome this may still bother some users. The aging population have shown a preference for remote sensors [7]. As described, they also struggle to detect certain types of non-typical falls, such as from beds. There is also the issue that some systems may be compromised if the user is not wearing the device correctly.
In order to overcome the technology’s shortcomings when recognising certain categories of falls, a machine learning approach could be taken. A machine learning wearable device has been already analysed [8] which monitors muscle activity with regards to maintaining balance, a similar approach could be taken to improve the detection of falls from a patient’s bed, by finding patterns between a body’s traits when lying in bed as opposed to on the floor. Another approach may be more customisable systems which are configured to the patient’s body proportions and bed heights in order to better inform the algorithms, or using localization techniques as described above in relation to [6].
An obstacle to the further development of wearable systems, particularly machine learning focused approaches, is a lack of training or test data. The vast majority of data is gathered from stunt people performing simulated falls, as opposed to weaker, older adults who are the systems intended users. Therefore, the test data is not entirely representative of the environment in which the systems will be used, due to differences in weight, speed of movement and involuntary reactions.
3.2 Vision Systems
Various vision-based systems have been developed to tackle the issue of fall detection, and the aforementioned [2], in 2013, stated that “vision based approaches in comparison to others are certainly the area to look forward to” despite the weaknesses found. These weaknesses included a “lack [of] flexibility”, as they tend to be more environment specific, and the need for manual initialisation.
A two-stage vision system using the Microsoft Kinect camera was proposed by the Centre for Eldercare and Rehabilitation at the University of Missouri, Columbia [9], which sought to be environmentally flexible in comparison to the predating systems. The system was evaluated using an “extensive real-world dataset” with the intention of a resulting higher performance when used in an actual home. The dataset is used to train a number of decision trees in the system. When placed into an independent living facility for evaluation, 454 falls were witnessed, the majority of which were planned stunt falls in order to tempt false positives from the system. When “The method was compared against five state-of-the-art algorithms...significantly better results were achieved”, however a recurring problem was that the low resolution at further distances provided by the Kinect meant that some falls went unnoticed. Unlike the wearable sensors described earlier in the report, the Kinect sensor was able to pick up horizontal falls, such as those from a sofa or bed.
Analysis
Accuracy for the detection of many types of falls is very high, including falls from horizontal positions. There is also the advantage that the patient must not remember to wear a device, and none of the functionality relies on the user; the camera is usually forgotten about. Many of the major flaws associated with camera methods, lie not with the algorithm or learning used, but the hardware choice in the fact that it is a vision-based system with limited range of sight; falls behind objects, such as those behind kitchen counters are not detected. Patients may be wary to carry out desired tasks, for example gardening, as they fear being away from the sensor, limiting the patient. As the sensor is observing all people in the visible area, there is a worry that more energetic actions carried out by people other than the patient, for example children dropping to the floor, maybe produce false positives. Lots of work in being carried out in the field of computer vision which will only contribute to the success of these systems [figure 2]. Systems are becoming increasingly better at distinguishing foreground from background, which will improve the detection accuracy of falls onto objects.
3.3 Ambience Approach
A recent innovation into the topic is a system titled “WiFall” and claims to be a “a truly unobtrusive fall detection system” looking for “correlations between different radio signal variations and activities by analysing radio propagation model” [10]. The system is trained under supervision using Random Forest (a learning method for categorizing data using multiple decision trees) to minimise the number of false Figure 2 alarms. The device can accurately detect the 4 main activities of walking, sitting, standing and falling, but so far only when there is one person within the monitored area. The approach shows lots of potential but is very limited by its aforementioned restriction.
Analysis
Due to the utilization of wireless networks, the system requires no wearable aspect, extra environmental configuration or hardware set up. However, the system cannot be put into use in isolation while it cannot work with multiple people in the room, it contradicts a lot of the benefits of living at home or with loved ones. There again is also the problem of the device not being useful once the patient is not at home. The creation of WiFall will “greatly contribute” [10] to the development of similar health care systems and shows good potential as an approach. The research being done into machine learning is only increasing with each year, which will overcome the flaws with the system. Unfortunately, further research into the system may be halted: “it is too costly to build the training database using learning methods since there are various combinations when considering different numbers, activities and locations of objects”.
SWOT Analysis of Fall Detection Systems, generally
Strengths:
Improve the independence of patients
Lessen the number of emergency hospital admissions
Give families peace of mind
Decrease the number of mortalities caused by unwitnessed falls
Weaknesses:
Certain category of falls are difficult to detect
Hardware can be expensive for families when using certain methods
Opportunities:
Demand is increasing
The older population are becoming more computer literate and may use the systems more independently
A global interest in machine learning will aid development
The systems are not region limited as they have no language or cultural barrier
Threats:
Regardless of the quality of the fall detection method, if the device cannot signal for help, it is made redundant. Services must have a good communication method for the detection systems to be worthwhile.
3. Conclusion
Overall, I believe that vision systems provide the most potential for development. With the continuing growth of research in computer vision [figure 2], the prospective advances in accuracy and environment flexibility present a lot of ambition for fall detection technologies. The lack of a wearable component makes the systems accessible to a variety of elderly people, regardless of their disability or illness. Although wearable systems can be used outside of the home, a lot of trips made by the elderly are with company, and there is no reason why an existing wearable solution may be worn when leaving, for extra precaution, if desired. With further research, multiple cameras are the obvious solution to overcome the issue of being out of sight. The development of these technologies is a great opportunity to improve the lives and safety of the increasing number of elderly and physically weakened people, as well as reducing the strain on caring and medical services.
References
[1] Public Health England, “Falls: applying All Our Health,” 14 June 2018. [Online]. Available: https://www.gov.uk/government/publications/falls-applying-all-our-health/falls-applying-all-ourhealth#facts-about-falls. [Accessed November 2018].
[2] L. S. L. S. Muhammad Mubashir, “A survey on fall detection: Principle and approaches,” Neurocomputing, vol. 100, pp. 144-152, 2013.
[3] K. D. J. R. J.Chen, “Wearable Sensors for Reliable Fall Detection,” in 2005 IEEE Engineering in Medicine and Biology 27th Annual Conference, Shanghai, 2006.
[4] H. J. Thomas Degen, “SPEEDY: a fall detector in a wrist watch,” in Seventh IEEE International Symposium on Wearable Computers, 2003. Proceedings., White Plains, NY, USA, 2003.
[5] L. R. M. A. Doughty Kevin, “The design of a practical and reliable fall detector for community and institutional telecare,” Journal of telemedicine and telecare, vol. 6, pp. 150-154, 2000.
[6] J. A. S. M. A. H. Qiang Li, “Accurate, Fast Fall Detection Using Gyroscopes and Accelerometer-Derived Posture Information,” in 2009 Sixth International Workshop on Wearable and Implantable Body Sensor Networks, Berkeley, CA, USA, 2009.
[7] M. R. George Demiris, “Older adults' attitudes towards and perceptions of "smart home" technologies: a pilot study.,” Medical informatics and the Internet in medicine, vol. 29, pp. 87-94, 2004.
[8] R. J. B. P. Hassan Ghasemzadeh, “A Body Sensor Network With Electromyogram,” IEEE TRANSACTIONS ON INFORMATION TECHNOLOGY IN BIOMEDICINE, vol. 14, no. 2, 2010.
[9] M. S. Erik E. Stone, “Fall Detection in Homes of Older Adults Using the Microsoft Kinect,” IEEE Journal of Biomedical and Health Informatics, vol. 19, no. 1, pp. 290-301, 2015.
[10] K. L. M. N. Yuxi Wang, “WiFall: Device-Free Fall Detection by Wireless Networks,” IEEE Transactions on Mobile Computing, vol. 16, no. 2, pp. 581-594, 2017.
Appendix
