Table 1 Comparison of context-aware systems
From: Challenges, issues and trends in fall detection systems
Article | Year | Basis | Features used for fall detection | Fall types | Subjects | Declared performance | Type of sensor | Elders Yes/No | Comments |
|---|---|---|---|---|---|---|---|---|---|
Lee et al.[20] | 2005 | Vision-based method for monitoring falls at home | State and geometrical orientation of the silhouette at time t, spatial orientation and speed of the centre of the silhouette | Fall lying down in a ‘stretched’ position and fall lying down in a ‘tucked’ position | 21 subjects (age 20–40) | SP: 80.5% | Camera | No | Personalized thresholds are established based on the height of the subjects |
SE: 93.9% | |||||||||
Miaou et al. [15] | 2006 | Customized fall detection system using omni-camera images | The ratio of people’s height and weight | Not specified | 20 subjects | With personal information: SP: 86% SE: 90% | Camera | No | Determining a proper threshold statistically for different ranges of height or weight alone does not improve the system performance |
Vishwakarma et al.[21] | 2007 | Automatic detection of human fall in video | Aspect ratio, horizontal and vertical gradient distribution of object in XY plane and fall angle | Sideways, forward, backward falls | 1 subject | SP: 100% SE: 100% | Camera | No | Both indoor and outdoor video containing different types of possible falls are taken |
Cucchiara et al.[19] | 2007 | A multi-camera vision system for detecting and tracking people and recognizing dangerous behaviours | Geometrical and colour features together with the projection of the silhouette’s shape on the x and y axes. | Not specified | Not specified | Difficulties with occlusions are reported | Camera | No | If a fall is suspected it delivers live video streams to clinicians in order to check the validity of a received alarm |
Fu et al.[16] | 2008 | Contrast vision system designed to detect accidental falls | Change in illumination | Backward, forward and sideways falls | 3 subjects | 3 possible scenarios evaluated with positive results | Contrast vision sensor | No | Instantaneous motion vectors are computed and fall hazards are immediately reported with low computational effort |
Hazelhoff et al.[17] | 2008 | Real-time vision system to detect fall incidents in unobserved home situations | The orientation of the main axis of the person and the ratio of the variances in horizontal and vertical direction Skin colour | Not specified | At least 2 subjects | SE: 100% when large occlusions are absent | Camera | No | The position of the head is taken into account in order to obtain a high robustness |
Anderson et al.[22] | 2009 | 3D representation of humans (voxels) using multiple cameras. Two levels of fuzzy logic determines first a state and then activities (f.i. a fall) | At low level: silhouettes from each camera, to build a set of voxels. At an intermediate level: centroid, height, major orientation of the body and similarity of the major orientation with the ground plane normal. | At least, falls forward, backwards, and to the side (with recovery, attempting to get back up, lying motionless) | Not specified | SE: 100% | Camera | No | The system can produce sentences like “the person is on-the-ground in the kitchen for a moderate amount of time” |
SP: 93.75% | |||||||||
Lie et al.[23] | 2010 | Vision fall detection system considering privacy issues | The ratio and difference of human body silhouette bounding box height and width | Not specified | 15 subjects (age 24–60) | Accuracy 84.44% | Camera | No | Activities are divided into three categories: standing posture, temporary posture and lying down posture |
Rimminen et al. [28] | 2010 | Fall-detection method using a floor sensor based on near-field imaging | Features related to the near-field imaging floor (the number of observations, the sum of magnitudes and dimensional features) | Backward to sit, backward to lateral, to supine, onto knees, arm protect, to prone, rotate right and left, right and left lateral | 10 subjects | SE: 91% | Near-field image sensor | No | The fall-detection performance is valid for multiple people in the same room |
SP: 91% | |||||||||
Tzeng et al.[25] | 2010 | A system that adjusts the detection sensitivity on a case-by-case basis to reduce unnecessary alarms | Floor pressure signal | Backward, forward and sideways falls | Not specified | SP: 96.7% | Pressure/ infrared sensors | No | The floor pressure sensor is combined with the infrared sensor |
Image features: standard deviation of vertical projection histogram, standard deviation of horizontal projection histogram, and aspect ratio | SE: 100% | ||||||||
Diraco et al.[24] | 2010 | An active vision system for the detection of falls and the recognition of postures for elderly homecare applications. | People’s silhouette and their centre-of-mass | Backward falls, forward falls, lateral falls | Not specified | SE: 80% | Camera | No | Information about the 3D position of the subject is combined with the detection of inactivity. |
SP: 97.3% | |||||||||
An approach for posture recognition is proposed | |||||||||
Rougier et al.[14] | 2011 | A vision system based on analyzing human shape deformation | Some edge points from the silhouette of the person | Forward falls, backward falls, falls when inappropriately sitting down, loss of balance | Not specified | Accuracy (falls and ADL correctly classified): 98% | Camera | No | The fall impact is an important feature to detect a fall, but the lack of movement after the fall is crucial for robustness |
Li et al.[29] | 2012 | Acoustic fall detection system | Acoustic signals sampled at 20 KHz | Backward, forward and sideways falls (balance, lose consciousness, trip, slip, reach chair, couch) | 3 subjects (2 female, 1 male, ages 30, 32, 46) | SE: 100% | Array of microphones | No | The source of the sound is located. |
SP: 97% | |||||||||
The performance of the acoustic detector is evaluated using simulated fall and nonfall sounds | |||||||||
Mastorakis et al.[18] | 2012 | Real-time fall detection system based on the Kinect sensor | The width, height and depth of the human posture, which define a 3D bounding box | Backward, forward and sideways falls | 8 subjects | All falls were accurately detected | Infrared sensor | No | The system requires no pre-knowledge of the scene and three parameters to operate; the width, height and depth of the subject |
Zhang et al.[27] | 2012 | Privacy Preserving Automatic Fall Detection | Deformation and person’s height | Fall from chair, fall from standing | 5 subjects | Accuracy 94% | RGBD cameras | No | The system can handle special cases such as light turning off (insufficient illumination) |