Falls are a major health hazard especially for the elderly and a serious hindrance for independent living. Falls are the major leading reason of injury related death for elderly people 79 years and older and second leading of injury related (unintentional) death for all ages. The demand for surveillance systems especially for fall detection is increased within the healthcare industry with the rapid growth of the population of the elderly in the world. Since falling becomes responsible for drastic physical- psychological outcomes, it has become very important to develop smart surveillance systems especially video monitored due to providing safe environments. Video Surveillance is an omnipresent topic when it comes to enhancing security and safety in the intelligent home environments. Multiple researches have proved that the medical consequence of a fall is highly contingent upon the response and rescue time. It must be decided at what level of abstraction to interpret the behaviour of the object and how to use the resulting data.
Thus, highly accurate automatic fall detection system is a significant part of the living environment for elderly to expedite and improve the medical care provided.
The quality of an individual's life is significantly affected by the levels of functional ability. Plenty of research has been done in this area to help in the development and assessment of functional ability. The modern development of cameras, sensors and computer technologies make such systems feasible. Such systems can not only increase the independent living ability of elderly by raising the confidence levels in a supportive care environment within public sector but also impact the manual labour in terms of the presence of nurses or support staff at all times. The fall incident should be detected as soon as possible so that the injured or affected can get the timely treatment.
Types of fall
In this section different kinds of falls are identified. Specifying different types of falls help towards understanding the existing approaches and algorithms. It also guides as well as contributes towards designing new algorithms.
Different scenarios are considered when identifying different kinds of falls: Falls from walking or standing, falls from standing on supports i.e. ladders etc., fall from sleeping or lying in the bed and falls from sitting on a chair.
There are some common characteristics among these falls as well as significant different characteristics. Noury et al [61] and Xinguo Yu [60] identified principles and methods used in existing fall detection approaches.
It is very interesting fact to notice that some characteristics of fall also exist in the normal activities e.g., crouch has a period of rapid reduction of head height etc.
Existing Approaches and Principles
In this section existing fall detection approaches and algorithms are categorised and explained in three different classes to build a hierarchy of fall detection methods. Later on different approaches under these categories are further discussed.
Existing fall detection methods can simply be divided into three categories: Wearable Devices, Camera based (Vision based) devices and ambience.
POSTURE DETECTION
SPATIO TEMPRAL
INACTIVITY
3D HEAD CHANGE ANALYSIS
POSTURE DETECTION
BODY SHAPE CHANGE
VISION
AUDIO & VIDEO
AUDIO
INACTIVITY DETECTION
TRI-AXIAL
ACCELEROMETER
WEARABLE SENSORS
AMBIENT/FUSION
FALL DETECTION
Figure 1 categorical flow chart of fall detection methodologies [60]
Wearable devices can further be sub divided into posture based and motion based devices. Ambience devices can be further sub divided into presence and posture based devices. While the camera (vision) based devices can be further sub divided into three classes as shape change, inactivity and 3D head motion.
Most of the existing approaches share the same general framework.
Figure 2 Framework for existing wearable sensors and ambient approaches [60]
Data acquisition varies from one sensor to multiple sensors and from one fixed camera to multiple cameras and moving cameras.
Wearable Device based Approach
The wearable device approach holds or uses garments with embedded sensors to detect the motion and the posture of the body of the object.
Accelerometery
Merryn et al [42] used an integrated approach of waist-mounted accelerometry. A fall is detected when negative acceleration is larger due to the change in orientation from upright to lying position occurring immediately. Bianchi et al [40] introduced a barometric pressure sensor, as a surrogate measure of altitude to improve upon existing accelerometry-based fall event detection techniques. The acceleration and air pressure data are recorded using a wearable device attached to the subject's waist and analysed offline. A heuristically trained decision tree classifier is used to label suspected falls. Estudillo-Valderrama et al [35] analysed results related to a fall detection system through data acquisition from multiple biomedical sensors, then processing and controlling the data with a personal server. The hardware and software design issues are clearly discussed when processing of bio-signals is involved during analysis. Tamura et al [36] developed a wearable airbag that incorporates fall detection by triggering airbag inflation when acceleration and angular velocity thresholds are achieved. The system design consists of accelerometer and a gyro sensor. Such fall detection system can be very useful especially at construction sites etc for reducing fall related injuries.
Chen et al [8] created the network by utilising small, non-invasive, low power motes to form a wireless low power sensor network. On-board device performs the sampling of acceleration sequentially, thus reduces the burden on the network. The dot product of acceleration vectors, from the orientation information, produces the angle of change before and after the fall event. The acceleration vectors are calculated using average of over one second window. Narayanan et al [19] developed a platform based on the "PreventaFall" ambulatory monitor (PFAM) and MiiLink data portal (MiiLink) is used to monitor accelerometric data. Wang et al [21] developed system that uses accelerometer placed on the head level. The algorithm distinguishes between falls and daily activities. Using backward integration of accelerations, the reference velocity is calculated. By using the reference velocity and predefine threshold, falls are distinguished from normal activities.
Fusion of Accelerometry & Posture Sensors
Luo et al [6] implemented a group of sensors on a belt which filter noisy components with a Gaussian filter and generate a three dimensional body motion model that can be related to various body postures and accelerometer's outputs by representing acceleration vector in 3D space. Kang et al [27] developed a wrist worn prototype integrated health monitoring device with tele-reporting function for emergency telemedicine that contains a fall detector. A two-axis accelerometer with a posture sensor is used in the fall detector. The measured bio-signals have limited fidelity because the wrist area has limited body contact. This shortcoming can be overcome with further development in the posture sensor.
Ghasemzadeh et al [37] used machine learning and statistical techniques to create a physiological monitoring system that collects acceleration and muscle activity signals and performs analysis on those signals during standing balance. The objective of this system is to assess the behaviour of the Electromyogram (EMG) signals to interpret the activity of postural control system in terms of balance control.
Inactivity with Accelerometry
Sixsmith et al [24] used a wearable system called (Smart Inactivity Monitor using Array-Based Detectors) SIMBAD that is made up of an array of relatively cheap infrared detectors. Target motion is analysed to detect characteristic dynamics of falls. Inactivity periods are also monitored and compared within the viewing field with a map of acceptable periods of inactivity in different locations. Ghasemzadeh et al in [38] implemented similar approach of inertial sensor nodes that constructs motion transcripts from biomedical signals and identifies movements by taking collaboration between the nodes into consideration. The system relies on motion transcripts that are built using mobile wearable inertial sensors.
Srinivasan et al [17] and Lee et al [19] both used motion sensors along with wireless accelerometer sensor module to monitor general presence or absence of motion.
Noury et al [3] designed a smart fall sensor for detection. The software application transmits the data remotely through the network as well as exploits data locally. The Boolean data is obtained for the position, vibrations and fall. The data is further analysed to determine the current state such as lying after a fall, sleeping, walking, etc.).
Tri-axial Accelerometry
Lai et al [41] combined several triaxial acceleration sensor devices for joint sensing of injured body parts, when an accidental fall occurs. The model transmits the information fed by the sensors distributed over various body parts. The model can determine the possible occurrence of fall accidents when the acceleration significantly exceeds the usual acceleration range. The impact acceleration and normal acceleration can be compared to determine the level of injury. Wu et al [30] developed the portable pre-impact fall detection system made up of inertial sensor and data logging unit. The interial sensor unit consists of accelerometers and tri-axial angular rate sensors. The inertial frame vertical velocity is the key variable that detects fall prior impact and is applied under threshold detection algorithm. Adoptive thresholding has been quite successful for the reduction of false positives.
Karantonis et al [7] designed the system that performs the vast majority of signal processing on-board the wearable unit using embedded intelligence. The triaxial accelerometer output is achieved from the portable unit containing the microcontroller embedded and the tracking information regarding user's motion is transmitted to a local receiver unit. Kangas et al [16] used acceleration thresholds to detect falls using triaxial accelerometric measurements taken at the waist, wrist, and head. The threshold values for different parameters are adjusted to optimal detection of falls.
Boissy et al [45] applied motion sensors on objects to derive impact magnitudes and trunk angle change. Motion sensors were placed on the front and side of the trunk along with three dimensional accelerometers. The above described features impact magnitudes while deceleration as hitting the ground and trunk angle change in relation to hitting the ground represented two separate events. The fall detection algorithm identified these two events as they are common to most falls. Wolf [46] et al followed a popular low cost approach of tri-axial accelerometer with wireless transceiver. The algorithm is very similar to other discussed accelerometric approaches in this review as data acquired from accelerometers is transmitted through wireless transceiver for further sophisticated analyses. The algorithm applied acceleration thresholds to detect falls.
Zhang's et al [49] has similarities to using the idea of wearable tri-axial accelerometer for fall detection but with the introduction of non-negative matrix factorization (NMF). The method used vertical axial of human body and acceleration sequences as input vectors. Vector decomposition is performed through NMF. Finally fall occurrence is determined via k-nearest neighbour algorithm. Zhang et al in [51] as opposed to [49] used interestingly a cell phone with a tri-axial accelerometer embedded. Data pre-processing is performed using 1-Class support vector machine (SVM) and wireless channel for internet connection. Precise classification is achieved through k-nearest neighbour (k-NN) algorithm and kernel fisher discriminant (KFD).
Posture Based
[52] Kaluza et al while working on the healthcare project designed to support independent elderly living called "The Confidence Project" presented a brief survey of different fall detection and activity recognition approaches and also discussed their suitable implications with respect to the on-going project. Falls along with abnormal behaviour detection is based on the ideology of reconstruction of object's posture. Small inexpensive wireless tags were placed on the places such as hips, ankles, knees, wrists, shoulders and elbows identifying them as significant places. The location of the tags is detected by the motion capture system. The posture is reconstructed in 3D plane after locating the tags. Acceleration thresholds along with velocity profiles are applied in the fall detection algorithm.
[56] Kangas et al carried out study with the aim to develop a new fall detector prototype. Waist worn tri-axial accelerometer, transceiver and microcontroller unit were used for data acquisition, transmission and processing. Sensitivity and specificity are also defined with respect to different fall detection algorithms.
Ambient Device based Approach
McKenna et al [1], from tracking data, automatically obtained spatial context models by using the combination of Bayesian Gaussian mixture estimation and minimum description length model order selection through Semantic regions (zones) of interest as Gaussian mixture components. Ceiling-mounted visual sensors are used to reduce occlusion. Thus resulting is the contextual model that detects the unusual inactivity an differentiates it from the normal activity. Tabar et al [2] presented Image sensing and vision-based reasoning for verification and further analysis of sensor transmitted events. A bridge like operation via wireless badge node is created between the user and the network. The badge node through event sensing functions detects falls. Besides fall detection it also creates a voice communication medium between the user and the Monitoring Control when the system detects an alert and calls the control. The monitoring control continuously tracks the approximate location of the user using signal strength measurements via the network nodes. A fusion of image sensing and network nodes is created for further analysis of field-of-view and user's status during fall detection.
Alwan et al [10] developed the working principle and the design of a floor vibration-based fall detector that is completely passive and unobtrusive. Detection of human falls is estimated by monitoring the floor vibration patterns. The principle is based on the vibration signature of the floor. The floor's vibration signature generated by the human fall is different from the normal activities like walking etc. A special piezoelectric sensor is used which is coupled to the floor surface. A battery powered pre-processing circuit alongside is used to analyse the vibration patterns. A binary fall signal is generated in the case of fall event.
Robinson et al in [23] proposed slip fall detection system using sliding linear investigative platform. The principle of classification between acceleration thresholds have been used to identify the true slip-falls detection. Low vibration translations are obtained by implementing the SLIP technology through an air bearing approach along with noncontact linear motor. Forces are measured such as tri-axial head accelerations and centre of pressure with shear force in terms of psychophysical response. SLIP-Falls vibrations are distinguished easily due to noticeable less vibration translations. The movement parameters require precise control and its advantages are discussed in terms of usefulness. Zigel et al [34] presented the concept of floor vibrations with sound sensing. Pattern recognition is applied to differentiate between fall and other events. Shock response spectrum is one of the key special features in classification. Human mimicking doll was used to simulate fall. The system is unique in detection of falls in critical cases such as object being unconscious or in a stress condition. The algorithm can be further developed with the calibration of the kind of floor.
Rimminen et al in [39] used a floor sensor based on near-field imaging. The shape, size, and magnitude of the patterns are used for classification. A collection of features is computed from the cluster of observations. The postural estimation is implemented using Bayesian filtering instead of the features being classified directly. The system had problems with test subjects falling onto their knees as it produced a pattern very similar to a standing person. Toreyin [48] et al used fusion of multitude of sound, vibration and passive infrared (PIR) sensors inside an intelligent environment equipped with above fusion elements. Wavelet based feature extraction is performed on data received from raw sensor outputs. Sound activity detector utilised the wavelet based features. Regular and unusual activities such as falls are used for training Hidden Markov Models (HMM). The process of fusion is applied to all outputs from sensors with classification to detect falls.
Zhuang et al [54] used different approach to that of Toreyin et al in [48] of using audio signal from a single far-field microphone. Gaussian Mixture model (GMM) super vector was created to model each fall as noise segment. The pairwise difference between audio segments is measured by the Euclidean distance. Kernel between GMM super vectors constituted towards the support vector machine employed for the classification of various types of noise and audio segments into falls. Doukas et al in [57] applied a mixture of accelerometric data with video streams in the algorithm. Wearable sensors transmit the motion data wirelessly. Precise classification is achieved from acquired data using support vector machines (SVM) to detect fall event. Finally video streams are transmitted from a context-aware server. The image sequences are coded accordingly to both patient and network status.
Nyan et al in [59] distinguished backward and sideway falls from normal activities using gyroscopes (angular rate sensors). The gyroscopes were securely placed on different positions such as underarm, waist etc. The angular rate is measured during normal activities and falls in lateral and sagittal body planes. High speed camera was used to capture video image sequences of motion for body configuration analysis in the event of fall. High speed cameras has frame rate of 250 frames per second. The fusion of high speed camera images and gyroscopes data is synchronised. Gyroscopes used the idea of acceleration thresholds to differentiate fall events from normal activities.
Camera (Vision) based Approach
Cameras are increasingly included these days in in-home assistive/care systems as they carry multiple advantages over sensors based devices including cost efficiency. Cameras can be used to detect multiple events simultaneously with less intrusion.
Khandoker et al [29] used support vector machines (SVM) for screening of balance impairments such as risks of fall in the elderly based on minimum foot clearance (MFC) principle which was used on the samples taken while walking on treadmill during the training. SVM model is based on the effectiveness of multi-scale analysis of a gait variable which is based on a wavelet in comparison to histogram plot analysis during feature extraction. There is a clear indication of better performance of the SVM model based on multiscale exponents (by wavelet analysis) in the results than the model based on MFC statistical features.
Spatiotemporal
Foroughi et al [4] discussed the method for detecting falls using a combination of Eigen space approach with integrated time motion images (ITMI). ITMI can be described as spatiotemporal database that contains motion information and time stamps of motion occurrence with emphasis at final action. Feature reduction is performed using Eigen space technique. Feature vectors obtained from feature reduction process are then fed to the Motion Recognition and Classification Neural Network classifier that can deal with motion data robustly.
Shi et al [33] used combination of computer vision and hardware but with a different approach. A mobile human airbag release system is designed for fall protection for the elderly. The system is designed using 3D MEMS accelerometers, gyroscopes, Micro Controller Unit and blue-tooth module. Object's motion information is recorded. A high speed camera is used for the analysis of falls. Gyro thresholding is applied to detect a lateral fall. The classification of falls is performed by using a support vector machine (SVM) filter during training. The real time fall detection system contains embedded digital signal processing system which is based on the SVM filter. Further development on SVM filter can lead to the algorithms be embedded into a real-time DSP.
Zhengming Fu et al [22] developed an address-event vision System that uses an asynchronous temporal contrast vision sensor. It extracts changing pixels from the background and reports temporal contrast (compared to set threshold), which is equivalent to change in image reflectance in the presence of constant lighting and finally an instantaneous motion vector computation reports fall events. The device requires a power socket nearby that makes the deployment of the system very simple. The motion detector protects the patient's privacy. No data is communicated until an emergency is detected.
Inactivity/Change of Shape
In shape change analysis algorithms as well as inactivity detection, Foroughi et al [5] together applied approximated ellipse around the human body for shape change. Projection histograms after segmentation are evaluated along with any temporal changes of head position is noted. Segmentation of moving objects is obtained initially, the next step involves extracting features by carrying out shape change analysis in the video sequence through approximated ellipse around the human body. Further analysis of projection histograms (both horizontal and vertical) and temporal changes of head position is carried out to extract feature vectors with optimised information. Extracted feature vectors are then fed to a MLP Neural Network similar to earlier approach of Foroughi in [4] for precise classification of motions and fall event. Miaou et al [9] designed detection system that uses an omni-camera called MapCam to capture images. The personal information of each individual such as weight, height and electronic health history is also considered in the image processing task. Object segmentation is performed by using methods like background subtraction. Noise reduction is applied during and after segmentation for accuracy. A bounding box like approach is used by creating a rectangle enclosing the object. The ratio of height to width of the object is calculated at each frame. The ratios are then analysed by considering last six consecutive image frames and resulting in altogether five ratio changes between two adjacent frames. The occurrence of fall becomes likely if the first three ratios are all greater than 1 and the last three ratios are all less than 1. System's decision of fall detection is based on the last two ratio changes with respect to the threshold. Each individual due to different body figures have different ratio changes between normal and fall states, the BMI (Body Mass Index) value is used to adjust the threshold. Therefore, the system is flexible enough to adjust the detection sensitivity on individual basis.
Tao et al [11] developed the detection system based on Miaou's et al [10] approach of using background subtraction (another approach based on shape change analysis algorithm) but with an addition of foreground extraction, extracting the aspect ratio (height over width) as one of the features for analysis, and an event-inference module which uses data parsing on image sequences. Simple two state machine in combination with falling motion inference is implemented. The two states are "standing/walking" and "falling down". Rougier et al [14]' approach based on a combination of motion history information (MHI) and human shape variation. The MHI contains an image information. The recent information of motion in an image sequence is represented by the pixel intensity such as the most recent movement during an action. Shape change analysis in combination with inactivity analysis is performed using approximated ellipse.
Lin et al [15] discussed fall incident detection in a compressed-domain. Object segmentation within compressed domain is applied for the extraction of moving objects using the combination of global motion estimation and local motion clustering. Three extracted features used are short time period range of fall occurrence, significant and rapid centroid change of falling human, and vertical projection histogram of falling human. Fleck et al [32] proposed a very unique idea of processing the stream at the point of sight and transmitting the processed stream to the control leaving no further processing to be done except the higher level abstraction. The system design consists of a distributed network that contains smart cameras. Geo-referenced tracking and activity recognition are performed simultaneously, embedded in each camera node. FPGA module and power PC processor are used for low level computations. The efficiency of the automated video analysis algorithms plays an important role towards the performance of the system. The proposed system could be further developed with self-diagnostic tools. Further improvements such as comprehensive processing and better decision making could be used as one of the major research directions for future development. This could significantly raise the level of robustness of algorithm and with respect to fault tolerance on hardware level.
[43] Ge wu et al implemented their method of automatic detection by uniquely identifying velocity profile features between normal and abnormal activities such as falls etc. The fall activities containing forward and backward falls from standing and tripping etc. Velocities such as Horizontal Vh and Vertical Vv were measured at different locations of the trunk. The trend of velocity increase showed an interesting pattern as it increased in one direction but did not in another. Two different characteristic patterns for falls were exhibited by Vh and Vv. Differentiating falls from normal activities during the descending phase of falls heavily depend on these characteristics i.e. change in magnitude and timing when the change in magnitude occurs of both Vh and Vv.
Vishwakarma et al in [44] followed an adaptive approach for the detection of moving objects by using background subtraction as well as marking bounding boxes. The described fall model is based on feature extraction analysis, detection and classification. Analysis contained features such as Horizontal and vertical gradients, aspect ratio and centroid angle to horizontal axis of bounding box. Falls were confirmed when the angle reaches a value less than 45 degrees. [47] Bromiley et al desinged the fall detector to monitor the image stream from the thermal detector for characteristic signals associated with falls. The analysis is focussed on measuring vertical velocities of the object using the coloured segmentation algorithm and identifying features in the pattern of velocities over time. These velocity estimates are then fed into neural network fall detector that identifies that identifies the characteristic patterns of velocities present during falls. Cucchiara et al in [53] instead applied multi-camera system for image streams processing. The processing included recognition of hazardous events and behaviours such as falls through tracking and detection. The cameras were partially overlapped. The cameras exchanged visual data during camera handover through a novel idea of warping "people's silhouette". The video server (multi-client, multi-threaded transcoding) transmitted sequences for further processing to confirm the validity of received data. The bandwidth usage is optimised through event-based transcoding and semantic methods. Anderson et al in [58] used multi-camera system like Cucchiara et al in [59] but used silhouettes to form a 3D model of the human object. The membership degree of the object is measured using fuzzy logic to a pre-determined number of states at each image. The fall detection method consisted of two levels. The first level deduced the number of states for object at each image. The second level dealt with linguistic summaries of object's states called as "Voxel Person". Further derivations were performed regarding the activity.
Posture
Cucchiara et al in [25] used computer vision to analyze human behaviours by classifying the posture of the monitored person and consequently detecting fall. Projection histograms are calculated and compared with the stored posture maps (training). The tracking also deals with occlusions. Accuracy levels achieved were up to 95%. Juang et al [32] used a posture classification approach based on a neural fuzzy network. standing, bending, sitting, and lying are the postures used for classification. After segmentation (background subtraction and extraction) projection histograms are used and discrete Fourier transform is applied. The design of classifier consists of a neural fuzzy network. The results could be improved with better segmentation such as better elimination of shadows and filtering illumination influence.
Thome et al [12] developed Hierarchical Hidden Markov Model (HHMM) with two layers for modelling motion. The first layer has two states, an upright standing pose and lying. Fall detection in term of sudden change has dedicated motion features from the first Layer. The understanding of 3D angles relationships and their image plane projection has been carefully observed. After performing an initial image metric rectification, theoretical properties are derived from binding the error angle for a standing posture during the image formation process. This very simply differentiates other poses as "non-standing" ones. Thus differentiating fall detection accurately from other actions such as walking or sitting. Computer vision systems typically consider cameras only for recording and capturing video images. Once transmission of the stream of images is completed, data processing is performed.
3D Head position Analysis
Rougier et al in [50] obtained streams from a monocular camera. This methodology of fall detection is based on 3D head trajectories and the idea of the object's head being visible in the image sequence and resulting into a large movement when a fall occurs. 3D ellipsoid is used for bounding around the head. The 3D ellipse is a projection of ellipse in 2D image plane. A particle filter extracts 3D head trajectory for tracking. 3D head trajectory also contains feature like 3D velocities which is applied for fall detection.
In 3D head motion based analysis, principle involving faster vertical motion than horizontal motion in a fall is used. Jansen et al [13] developed the method that uses information extracted from images obtained using three dimensional visual approach in combination with a context model. The contextual model interprets fall occurrence differently. It depends on the time, location and duration of the fall event.
Brief Comparison of Characteristics
Approach
Cost
Accuracy
Setup
Robust
Wearable Devices
Cheap to Medim
High/specific
Easy
No
Ambient
Cheap to Medim
Higher / Specific
Easy / Medium
No
Vision based
Medium
Higher / Non Specific
Easy / Medium
Yes
Different Scenarios for Evaluation
Noury et al [61] proposed different scenarios for detectors. The following table is taken from Noury et al in [61].
Category
Name
Outcome
Forward Fall
On the Knees
Positive
With Forward Arm Protection
Positive
Ending lying flat
Positive
With rotation, ending in the lateral right
position
Positive
With rotation, ending in the lateral to the
left position
Positive
With recovery
Negative
Backward Fall
Ending sitting
Positive
Ending lying
Positive
Ending in lateral position
Positive
With recovery
Negative
Lateral Fall to the Right
Ending lying flat
Positive
With recovery
Negative
Lateral Fall to the Left
Ending lying flat
Positive
With recovery
Negative
Syncope
Vertical slipping against a wall finishing
in sitting position
Negative
Neutral
To sit down on a chair then to stand up
(consider the height of the chair)
Negative
To lie down on the bed then to rise up
Negative
Walk a few meters
Negative
To bend down, catch something on the
floor, then to rise up
Negative
To cough or sneeze
Negative
Table 1 Scenarios for Fall Detection [61]
Conclusion and Future Work
The ideal fall detection system should exhibit both sensitivity and specificity. The existing approaches have not comprehensively satisfied the accuracy as well as robustness of fall detection system. Although the existing approaches do provide a framework to further develop techniques as well as modify the existing algorithms to achieve better performance.
Vision based approach is certainly the area to look forward to. Most of the existing vision based approaches lack flexibility. These approaches are often case specific dependent on different scenarios. This requires a need for a reliable and robust generic fall detection algorithm.
This also puts stress on the availability of data sets of falls for training as well as the evaluation. A comprehensive data set containing different scenarios of falls with different camera angles and with both static and moving cameras should be publicly available for the scientific community for the development and research purposes.
Continuous surveillance through vision/camera based systems also introduces some ethical issues concerning the respect of intimacy and privacy and also the risk of dependency of the subject on the technology. A common definition of a fall and of fall detection system would certainly benefit the research community as well as the healthcare industry for the evaluation of the fall detection systems.
