Fall is a commonplace event which might happen from early childhood to old age resulting on loss of dignity to injuries or it may be even fatal for older population. In 1987 Kellogg international working group defined elderly fall as "unintentionally coming to the ground or some lower level and other than as a consequence of sustaining a violent blow, loss of consciousness, sudden onset of paralysis as in stroke or an epileptic seizure".
Falls might happen on steps, stairs, slopes, and from height with different cause and consequences, together leading to a major source of injury, imposing a substantial social and economic burden to the society. Falls can be categorised in several ways according to the circumstances, age and prevention. Here we differentiate falls by research communities who most directly addressing to the problem.
Falls among older population which is a priority concern for healthcare, medical and gerontology professionals which received the most research attention. Research work is increasing in this area nowadays due to ageing community around the world.
Falls in workplace were occupational health and safety professionals work with less research work on this area.
Elderly population:
Elderly population and their diseases is getting much attention on research work nowadays, Singapore department of statistics stated that "The proportion of Singapore residents aged 65 and above increased from 6.8% of the resident population in 1998 to 8.7% in 2008" (singstat) and this projection was expected to rise around 18.7% in 2030@.
United Nations also estimated that by the year of 2100, 28.1% of the world population will be above 65 years of age which accounted for 10% of world population in the year 2000(1). So each and every nation is to develop health monitoring and support systems for elderly persons, which should be more effective and efficient in its nature.
Due to increase in elderly population there is a decline on working group which in turn gives less care for elderly population from the working group. A Singapore survey done in the year 2000 shows that 6.6% of older persons aged 65 and above lived alone while 14% lived with their spouses and 74% lived with their children, whom may be working in this situation there is a need for universally accepted elderly monitoring device or system which should be accepted by the elderly and their care givers.
Falls in elderly
One among three elderly persons is likely to fall more than one times a year. A survey on 1999 states that fall is the second most common cause for emergency among elderly peoples in Singapore, another study states that 85.3% of injuries in elderly are due to falls!. Several studies had stated that fall increases likelihood of long term hospitalisation or under health care institution.
In elderly persons it is hard to maintain postural stability due to poor interaction of sensori, motor, and integrative systems. Some other reasons like impaired vision, strength, sensation, and reaction time also leads to poor balance in elderly population. When a fall occurs among elderly peoples who were living lonely it will be gone unnoticed for several hours, days, or even several weeks. So, there arises a situation to monitor elderly falls independent of either they live in residential houses or nursing homes and to alert their relatives and care givers about the falls. Researchers are working to formulate system, methods and devices to detect these falls and alert their care givers.
Fall detection:
There are a number commercial devices used to alert elderly care givers which were worn like a medallion, wrist watch or wall mounted switches which are designed as manual alert system which were failed on alerting. After a fall elderly peoples are unable to activate these devices due to their physical inability or memory loss and they might forget to wear these devices on their daily activities.
Researchers has worked on a number of devices and methods for fall detection which are like sensors, accelerometers, gyroscope, video monitors, floor vibration devices and sound alert systems and formulated number of methods and algorithms like manual and machine learning methods each and every system and method has its advantages and disadvantages which will be discussed in the following sections and future research work will be discussed.
A fall can be differentiated into four types namely forward fall, backward fall and two sideways falls. A fall detector should be designed in a way it should differentiate activities of daily (ADL) a falls. Some devices in the market can easily differentiate falls from low intensity daily activities but most of the devices are unable to detect the fall from high intensity activities like leaning on a chair. This is the area were researchers are fine tuning their devices
There are a number of algorithms and methods used for fall detection like analytical methods, machine learning methods.
Methods
There are a number of devices and methods used for falls detection which will be described briefly in the following.
2.1 Fall detecting Devices
2.1.1 Social alarms
Social alarms are commercially available devices in the market. These are the devices like alarms for care givers which might be worn by the patients like a wrist watch, chain, wall mounted switches which are manually operated on an emergency situation and patients can talk to their care givers through a phone attached to these devices. These are the major failure devices which are not operated by the patients on an emergency situation due inability to operate these devices or they might forget to wear these devices.
These devices are simple and low cost. These are not effective when the falls associated with consciousness, trauma, pain patients are unable to activate these devices. These social alarms also have problems by giving falls alarm when patients remove these devices when they go for shower which leads to discriminating these products both by the patients and their care givers.
Medical emergency products
2.1.2 Accelerometers
Accelerometers are devices which are used to determine the magnitude and direction of acceleration in a single axis or multiple axes. In research tri axial accelerometers are mostly used, by detecting the acceleration with earth's gravity one can also compute angle with earth's gravity. Simple methodology for fall detection is using a tri axial accelerometers with threshold algorithms, when a threshold acceleration limit is reached it will raise alarm to alert either patients relatives or care givers, these devices work with an accuracy of more than 80%.
Francis E.H. Tay et al from National University of Singapore developed a smart shirt in 2005 and placed a single axis accelerometer on the shoulder part of that shirt and defined a threshold limit of 1.2V or 4.8 g to differentiate a fall from normal activity but it was not tested for its sensitivity and specificity (39).
Nyan et al from national university of Singapore, have contributed a lot for falls detection they had used a 3 axis MEMS accelerometer which were very compact to use and it can be used in a garment and in shoulder area to detect patients ADL and falls and they had succeeded above 95% both in sensitivity and specificity of falls (9, 10, 11)
Bourke et al. had used a tri axial accelerometer to frame a fall detection algorithm, using middle aged subjects he has done a number of simulation of falls and elderly persons are used to determine their activities of daily life (ADL). Those tri axial accelerometers are located at the trunk to differentiate falls from ADL (2)
Pei kuang cho et al. has used acceleration cross product (AC) related methods, proposed and examined by this study to seek solutions for detecting falls with less motion-evoked false alarms. A set of tri-axial acceleration data is collected during simulated falls, posture transfers and dynamic activities by wireless sensors for making methodological comparisons. The performance of fall detection is evaluated in aspects of parameter comparison, threshold selection, sensor placement and post-fall posture (PP) recruitment. By parameter comparison, AC leads to a larger area under the receiver operating characteristic (ROC) curve than acceleration magnitude (AM) (4).
Diaz et al, evaluated intelligent accelerometer unit (IAU) which can be used on elderly persons to monitor and differentiate their ADL and falls which are framed for a tele- health care system(5) to monitor a patients activities.
Tong zhang et al has embedded tri axial accelerometer in cell phone. In order to increase the feasibility and effectiveness they had used accelerometers as wearable devices it will be placed on a pocket or hanged around neck and they had tested this device with both elder and younger population (38). If these cell phone mounted accelerometers are designed like a body mounted devices they will be more efficient.
Tong Zhang et al had designed a waist mounted tri axial accelerometers and used support vector machine (SVM) algorithm, which can learn acceleration in each direction and change in acceleration and they had succeeded with 96.7% accuracy on a laboratory based experiments (28).
Tong zhang et al also designed a tri axial accelerometer which were used on a belt and wear on abdomen and used a Non negative matrix factorization (NMF) algorithm which differentiated fall from daily activities. But this algorithm had a high false alarm with high intensity daily activities. NMF algorithm had a overall efficiency of 95%.
Charalampos et al designed a tri axial accelerometer which uses support vector machine algorithm to classify patient's movement activity and falls. This device monitors user movements which give movements in terms of acceleration in three axes and support vector machine provides fall detection with 98.2% accuracy. When a fall is detected patients video image coding is also transmitted through a tele-monitoring device.
Thomas degen et al designed "SPEEDY" a wrist watch accelerometer which is an easy wearable device on daily activities since wrist is the most difficult place to detect fall they had designed their algorithm according to it and at last they had concluded if it is wearable on other parts of body "SPEEDY" will be more accurate.(14).
Tapia et al. [40] presented a real-time algorithm for automatic recognition both physical activities and their intensities, using five wireless accelerometers and a wireless heart rate monitor. The accelerometers were placed at shoulder, wrist, hip, upper part of the thigh and ankle. The features, e.g., FFT peaks, variance, energy, correlation coefficients, were extracted from time and frequency domains using a predefined window size on the signal. The activities were classified into three groups: 1. Postures like standing, sitting etc., 2. Activities like walking, cycling etc. and 3. Activities like running, using stairs etc. For these three classes they obtained the recognition accuracy of 94.6 % using subject-dependent training.
Researchers give more attention for sensor placement in the body. Head is the most efficient place for sensor placement in accelerometer methods but it has some limitation like patient and care givers acceptability. Waist worn accelerometers are also more efficient than any other devices or methods on experimental setup.
Accelerometer either it is a wrist worn, waist worn, garment worn, cell phone placed or thigh and shoulder worn it leads a drawback on patients comfort level and some patients might forget to wear these devices on their daily activities like taking a shower. False alarms are unavoidable by these devices on some activities like remove and place those devices for a shower etc. This leads to discriminate these devices both by the patients and their care givers.
2.1.3 Gyroscope
Gyroscope is used to measure orientation. Gyroscope consists of a spinning wheel whose axle is free to move on any orientation. It can measure orientation on single axis or multiple axes. By equipping an object like accelerometer with gyroscope we can measure orientation along multiple axes. Gyroscope is used to measure objects orientation and its change in orientation which is used to compute angular velocity.
Efficiency to differentiate high intensity tasks from falls is very low on accelerometers, so researchers like qiang li et al and bourke et al used gyroscopes to differentiate falls in order to increase the efficiency of fall detection.
Qiang li et al used a sensor which consists of both tri axial accelerometer and tri axial gyroscope which are placed on chest and leg on subject's body and they have measured angular velocity and linear acceleration for four different activities and framed a threshold limit to differentiate falls from ADL. In this algorithm they had reduced both false negative and false positives which were faced on accelerometers. But their algorithm failed to differentiate activities like jumping into the bed, falling against wall with seated posture from falls.
Bourke et Al. framed a threshold based algorithm using bi axial gyroscope sensor to differentiate falls from ADL. Young volunteers performed simulated falls on a mat while elderly peoples performed ADL. A gyroscope is mounted on the trunk to determine angular acceleration, angular velocity and trunk angle a threshold limit is obtained to differentiate falls from ADL. Tests were carried for 480 simulated movements results showed that fall is differentiated from ADL with 100% accuracy (3).
Since bourke et al achieved 100 % accuracy and 97.5% specificity on a bi axial gyroscope in test environment it should be applicable in a real life situation.
2.1.4 Visual fall detection
Since video cameras and surveillance camera is common nowadays researcher used cameras to monitor patients it may be either analytical method or machine learning methods. Nursing homes equipped with surveillance camera with fewer patients are monitored by analytical method personals are used to monitor patient's activities or falls and can activate alarm.
Viswakarma et al used a video approach for fall detection first they subtracted background from the video using Gaussian model mixture and they extracted remaining features from the objects as bounding boxes such as aspect ratio, horizontal and vertical gradient values and fall angle. A model is build which consists of two steps, first step is fall detection for which they used aspect ratio and objects horizontal and vertical gradients next step is fall confirmation which uses fall angle with respect to horizontal axis of its bounding boxes. This algorithm detects a fall when there is a change in aspect ratio, vertical and horizontal axis then a fall is confirmed when a fall angle less than 45 degree with respect to horizontal axis of its bounding boxes. This model is tested with 45 video clips on which they achieved 95% accuracy on single object fall detection 64% accuracy on multiple fall detection.
Xinguo yu et al framed an algorithm for fall detection in which three steps torso ellipse acquisition, shape state detection in which they used three shape states namely standing, bending and lying then state change pattern then detect the fall.
Lykele et al used two cameras for fall detection both are used on perpendicular for tracking a single person images are taken by those two synchronised camera foreground is obtained from background subtraction, since it is used for single person detection objects are tracked by both the cameras then both camera fuses human objects then a fall is detected by Gaussian classifier after a fall is detected head position is tracked and it is used for false rejection. Experimental results show that they achieved 85% accuracy in real time speed.
Bernd schulze et al used a single video camera mounted on the ceiling and it cannot be used for monitoring other activities it is mainly used to detect falls. Single camera is used to monitor a single room they had divided the room into three zones according to users activity which are standard zone, entry/exit zone and low activity zone. Fall detection code consists of three parts namely image download or pre processing, motion analysis and fall detection. Fall detection algorithm is written such that it will monitor motion of the person according to the zone if there is no motion on the standard zone it will raise an alarm. Drawbacks are like It is intended only for a single person, single room and fall detection monitoring, this method does not use any object recognition and shadow subtraction which can lead to high false alarm.
Bart jansen et al used a 3D camera for fall detection in an experimental setup room the room is designed in order to prevent occlusions for 3D camera they used a 7 point torso and given algorithm for activity and inactivity detection. Since this is a preliminary test further research work will be carried to answer those difficulties faced by this experiment.
Jared Williams et al framed an algorithm for single camera fall detection in which they subtracted the background and obtained shadow detection through fall angle, aspect ratio and vertical projection histograms results shows that shadow detection is 85% accurate on side view and 78% accurate on front view.
Charalampos doukas et al used both accelerometer and video for fall detection they used SVM algorithm both for accelerometer and video detection first a fall is detected by accelerometer later it is confirmed with video images. Test results shows 98% efficiency of this system.
2.1.5 Fall detection through Floor vibration and sound
Litvak et al, (8) on august 2008 framed a solution based on vibration and acoustic sensing, they used a sensor which consists of accelerometer and microphone which acquires sound and vibrations from the floor and it is monitored on a pc using energy based event detection algorithm. Falls, daily activities and inhuman activities are recognised through this algorithm an alarm is activated when it detects a human fall which has proved a sensitivity of 95% and specificity of 95% which also can be monitored in a remote clinic.
Litvak et al on December 2009 used their sensors which consists of accelerometer and microphone and with improved fall detection algorithm conducted tests with dummy on the basis of machine learning algorithms. Test showed improved results on identifying and classifying human falls from an object fall with 97.5% sensitivity and 98.6% specificity. One of the advantages of this test is conducted on a concrete floor which is a common floor. But the algorithm should be trained with lots of real life impact and sounds, which also has some limitation like it could not classify low impact human falls from object falls.
Majd elwan et al designed a passive floor vibration fall detector which uses a special piezo electric sensor coupled to the floor surface with mass and spring. The device has a battery powered pre-processing electronics which transmits fall alarm to communication gateway. Human fall signals are stored on the device which will detect human falls and send as a message to a cell phone. This device can detect a fall up to 20 feet range and should kept 2 feet away from walls in order to avoid vibration detection from the neighbouring rooms test was carried out with two simulation using a dummy one fall from wheel chair and another from a bed and they had reached a 100% sensitivity and specificity. The device has lot of limitation for its accuracy.
2.1.6 Asynchronous temporal contrast (ATC) vision sensor
A temporal contrast sensor is a device which extracts the changing pixels from background and reports temporal contrast which is equivalent to image reflectance when the lighting is constant. zhengming fu et al used a temporal contrast vision sensor which are placed at a distance of 3 meter and with height of 0.8 meter. These imaging sensors streams series of events with time stamped address events from the vision and it is send back to the pc. An event is computed with address and time and it is evaluated on a PC.
