Table 2 Fall detection techniques in the context-aware studies
From: Challenges, issues and trends in fall detection systems
| Â | Fall detection method | ||||
|---|---|---|---|---|---|
1ststep | 2ndstep | 3rdstep | 4thstep | 5thstep | |
Lee et al.[20] | Adaptive background subtraction to detect the object of interest | Image processing using a connective-component labelling technique, with the end product being a ‘blob’ or silhouette | Feature extraction | Determination of the threshold values for each of the features based on the height of the users |  |
Miaou et al.[15] | Background subtraction to detect the objects. | Image processing: erosion and dilatation, connected component labelling technique | Feature extraction (height and width of object’s silhouettes) | Simple threshold-based decision algorithm for fall detection |  |
Vishwakarma et al.[21] | Patient detection (adaptive background subtraction method using Gaussian mixture model) | Feature extraction | Fall detection using aspect ratio and pixel's gradient distribution and applying rule-based decisions | Fall confirmation using the fall angle and applying rule-based decisions | Â |
Cucchiara et al.[19] | Extraction of moving objects using background suppression with selective and adaptive update | Tracking algorithm: A probabilistic and appearance-based tracking | Classification as people of tracks that satisfy some geometrical and colour constraints | Posture classifier based on the projection histograms computed over the temporal probabilistic maps obtained by the tracker | Hidden Markov Models formulation is adopted to classify the posture |
Fu et al.[16] | Extraction of changing pixels (motion events) from the background | A lightweight algorithm computes the instantaneous motion vectors | Fall events are reported using the temporal average of the motion events | Â | Â |
Hazelhoff et al.[17] | Object segmentation: (background subtraction and connection of information components) | Object tracking: the tracker can mark objects as non-human, which are identified based on size and absence of both motion and a head region | PCA-based feature extraction: the direction of the principal component and the variance ratio are extracted | Fall detection: using a multi-frame Gaussian classifier | Head tracking using skin-colour model to confirm the fall |
Anderson et al. [22] | Silhouette extraction from each camera. Then, a 3D representation of the body is constructed | Extraction of centroid, height, major orientation of the body and similarity of the major orientation with the ground plane normal | Human state inferred using fuzzy logic (3 states: upright, on-the-ground and in-between) | Information in sequences of states is reduced by linguistic summarization to produce human readable sentences | Fall detected by a second level of fuzzy logic, taking inputs from a single summary: average state, time duration, speed, oscillation, etc. |
Lie et al.[23] | Human body identification using frame differencing approach | Image processing: mean filter to make the image more smooth, thresholding to obtain a binary image, connected component labelling | Features extraction and reduction of upper limb activities effect | k-nearest neighbour classifier for human body postures classification | Fall event detection flow: the decision of a fall incident is determined by the event transition and time difference between events |
Rimminen et al.[28] | Estimate the position of the subject using the near-field image sensor observations | Tracking (Kalman filter) and multi-target tracking (Rao-Blackwellized Monte Carlo data association algorithm) | Features extraction related to the NFI floor | Modelling of the state evolution as a two-state Markov chain (falling, getting up) | Pose estimation using Bayesian filtering. It combines the prior model with information from the features |
Tzeng et al.[25] | Fall suspection: Thresholding of the floor pressure signal | If the floor preassure exceeds a given threshold: Image capture | Background subtraction through an image thresholding. Objects labelling and expansion (morphological operations) | Image features extraction | Combination of the floor pressure signal and image features to report on a fall |
Diraco et al.[24] | Camera calibration | Background modelling using Mixture of Gaussians method | Moving regions detection (Bayesian segmentation) and segmented blobs refining (morphological operations and connected components) | Fall suspection: The distance of the centroid from the floor plane is lower than a prefixed value | Fall confirmation if an unchangeable situation persists for at least 4Â seconds |
Rougier et al.[14] | Silhouette detection (foreground segmentation method) and edge points extraction (Canny edge detector) | Silhouette edge points matching through the video sequence | Shape analysis using the mean matching cost and the full Procrustes distance | Fall classification: Gaussian mixture model, based on shape deformation during the fall and the subsequent lack of movement | Â |
Li et al.[29] | Locate the position of the sound source | Beamforming to enhance the sound signal using the estimated source position | Mel-frequency cepstral coefficients features are extracted from the sound signal | A nearest neighbour classifier determines if the sound is from a fall | Â |
Mastorakis et al.[18] | Feature extraction: width, height and depth of the human posture | Obtaining of the velocities of height and the composite vector of width and depth | When both velocities exceed particular thresholds fall initiation is detected | Inactivity detection: a fall is detected if the height velocity is less than a certain threshold | Â |
Zhang et al.[27] | Kinematic Model Based Feature Extraction from Depth Channel | Person tracking by background subtraction | Histogram represented features | Hierarchy Support Vector Machine classification | Â |