An Integrated Vision-Based Approach For Efficient

This article has been accepted for publication in a future issue of this journal, but has not been
fully edited. Content may change prior to final publication. Citation information: DOI
10.1109/ACCESS.2019.2936320, IEEE Access
Date of publication xxxx 00, 0000, date of current version xxxx 00, 0000.
Digital Object Identifier 10.1109/ACCESS.2017.DOI
An integrated vision-based approach for

efficient human fall detection in a home
environment
FOUZI HARROU1 , NABIL ZERROUKI2 , (MEMBER, IEEE), YING SUN1 AND AMRANE
HOUACINE2 .
2
King Abdullah University of Science and Technology (KAUST), Computer, Electrical and Mathematical Sciences and Engineering (CEMSE) Division, Thuwal,
23955-6900, Saudi Arabia.)
1
University of Sciences and Technology Houari Boumédienne (USTHB), LCPTS, Faculty of Electronics and Computer Science, Algiers, Algeria.
Corresponding authors: Fouzi Harrou (e-mail: [email protected])
This work is supported by the King Abdullah University of Science and Technology (KAUST) Office of Sponsored Research (OSR) under
Award No: OSR-2019-CRG7-3800.
ABSTRACT Falls are an important healthcare problem for vulnerable persons like seniors. Response to
potential emergencies can be fastened timely detection and classification of falls. This paper addresses the
detection of human falls using relevant pixel-based features reflecting variations in body shape. Specifically,
the human body is divided into five partitions that correspond to five partial occupancy areas. For each
frame, area ratios are calculated and used as input data for fall detection and classification. First, the
detection of falls is addressed from a statistical point of view as an anomaly detection problem. Towards
this end, an integrated approach merging a detection step with a classification step is proposed for enabling
efficient human fall detection in a home environment. In this regard, an effective fall detection approach
using generalized likelihood ratio (GLR) scheme is designed. However, a GLR scheme cannot discriminate
between true falls and like-fall events, such as lying down. To mitigate this limitation, the support vector
machine algorithm has been successfully applied on features of the detected fall to recognize the type of fall.
Tests on two publicly available datasets show the effectiveness of the proposed approach to appropriately
detecting and identifying falls. Compared with the neural network, k-nearest neighbor, decision tree and
naïve Bayes procedures, the two steps approach achieved better detection performance.
INDEX TERMS Smart home; human fall; fall detection; classification; machine learning algorithms.
I. INTRODUCTION human health and safety [3] has enabled to the develop-
ALLS are an important health problem for seniors as ment of many fall detection techniques [4] that can be
F they might generate serious injury [1], [2]. Efficient
detection and classification of falls can be very helpful to
distinguished into two essential categories: non-computer-
vision-based and computer-vision-based methods [4]–[7].
achieve a quicker response to potential emergencies. Once Non-computer-vision approaches are usually based on data
a fall is detected, the classification module must discriminate acquired by sensors. These techniques utilize sound, vibra-
between real falls and fall-like activities (e.g., lying down). tions and human motion in detecting falls [8], [9]. Actually,
Detection needs to be immediate and classification needs to numerous daily devices like smartphones and smartwatches
be accurate. incorporate small sensors [10], [11].
To achieve reliable fall detection, fall prevention, interven- Other researchers use information learned from images
tion, and safety become increasingly important and strenuous and videos in detecting falls [12]. These mechanisms were
efforts have been invested throughout the last two decades. introduced based on a single camera, multiple cameras [13],
Various projects (e.g., the European profound project) were and omnidirectional ones [14] and stereo-pair cameras [15].
initiated in worldwide to respond to this imminent need This approach presents some benefits over a sensor-based
to improve the elder’s quality of life. Over the last few approach, in particular, no human intervention is needed and
decades, increasing attention in human fall detection for no wearing of sensors is required. In comparison to wearable
VOLUME 4, 2016 1
This work is licensed under a Creative Commons Attribution 4.0 License. For more information, see https://creativecommons.org/licenses/by/4.0/.
This article has been accepted for publication in a future issue of this journal, but has not been fully edited. Content may change prior to final publication. Citation information: DOI
F. Harrou et al.: An integrated vision-based mechanism for efficient human fall detection in a home environment
sensors, the camera provides more complete and detailed this work, an efficient computer vision approach is proposed
information about the supervised person (e.g., posture, po- to address the fall detection and classification problem.
sitioning, and location), as well as their environment. Thus,
growing demand for vision-based fall detection systems has This paper introduces an effective and efficient human
been recorded from health centers, hospitals, and retirement fall detection mechanism based on the generalized likeli-
homes [16], [17]. Furthermore, many research in related hood ratio (GLR) approach, which is an efficient anomaly
fields has shown that vision-based remote health monitoring detection technique [28]. The GLR detector is suitable to
is essential to monitor non-critical patients at home rather achieve automated detection of falls due to its capability to
than in hospital whilst reducing strain on hospital resources. small changes [28], [29]. However, the principal shortcoming
Since video surveillance systems enable efficient analysis of of the GLR detector is its incapability to separate real falls
the behavior of the supervised person (even in the normal from behaviors similar to a fall like stretching out. Then, to
state), such as taking medication and sleep schedules, it mitigate this limitation, the greater discrimination capacity of
facilitates to elderly people to live independently at home for SVM is exploited. Notably, the SVM classifier is frequently
longer [17], [18]. In the last past few years, several interna- used for tackling the classification of linear and nonlinear
tional projects such as BIOTELEKINESY are implemented features by employing nonlinear kernels. The main benefit of
using camera-based fall detection mechanisms [19]. To mit- SVMs consists in their capacity to suitably uncover important
igate the problem of stakeholder’ privacy, image acquisition, patterns in the data by transforming problems into higher
and data treatment should be consistent with ethical concerns. dimensions using kernel functions, enabling a non-linear
To do so, only the processed images are used for fall detection relationship to appear approximately linear [30]. Here, the
and the videos of the supervised senior are not recorded [19]. desirable properties of the GLR chart and the SVM classifi-
In [20], Ozcan et al. introduced a systematic fall detection cation system are exploited for enabling efficient human fall
procedure via wearable cameras using relative entropy met- detection in a home environment In this work, the classifica-
ric. Rougier et al [21] designed a fall detection technique by tion stage is performed only when a potential fall is identified
merging the variation of human shape and motion history. by the GLR approach. Therefore, the SVM classification is
However, this approach uses a threshold manually fixed to fed with a reduced number of sequences, which enhance
discriminate non-fall from fall activities. Also, a high false accuracy. Accordingly, by merging the GLR fall detection
alarm rate resulted due to miss-classification of several non- procedure with the SVM classifier, GLR-SVM, fall detection
fall activities such as fast sitting activities. Miaou et al. [14], accuracy is significantly improved compared to that ob-
[22] proposed a fall detection algorithm using a MapCam tained via the conventional approaches. Tests on two publicly
(omnidirectional camera). Specifically, this approach uses the available datasets show that the proposed GLR-SVM has a
rule-based algorithm. This approach achieved 91% success- good capacity for enabling efficient fall detection. Compared
ful fall detection rate. In [23] a reconstructed 3-dimensional with the conventional machine learning techniques namely
human silhouette was used to determine whether a person neural network, k-nearest neighbor, decision tree, and naïve
fell. The decision in this mechanism is based on the volume Bayes procedures, the GLR-SVM mechanism achieves better
distribution along the vertical axis, an alarm is reported if the detection results.
majority of this distribution was abruptly close to the floor.
The contribution of this study is threefold: (1) relevant
The method in [23] requires multiple cameras and a graphics
pixel-based features from images have been introduced and
processing unit for processing. The method in [24] used an
adopted for fall detection. (2) The GLR is used during the
approximated ellipses that closely encapsulate the contours
detection phase to discriminate fall from fall-free features. (3)
of the human body and temporal changes of head position
The SVM algorithm is used to bypass fall-like actions. The
to detect human fall; falls were classified by neural network
rest of this paper is arranged as follows. Section II, we give
algorithm. Other works used a depth camera and sensor data
a brief description of segmentation and feature extraction.
to detect potential falls by support vector machine (SVM)
In Section III, GLR approach and its use in fall detection
algorithm [25], [26]. In [27], an image-based fall detection
are briefly introduced, and the proposed approach for fall
approach has been introduced to detect falls in the elderly
detection and classification are outlined. In Section IV, the
based on videos from surveillance systems or webcams. This
performance of the proposed method is assessed. Lastly,
approach is performed into two steps: after identifying people
Section V concludes this work.
in a video frame using an object detection algorithm, then
a posture recognition method is applied to track the status
of the supervised person by verifying the relative positions II. A VISION-BASED FALL DETECTION FRAMEWORK
of the chair and the people [27]. An alarm is given when The proposed fall-detection and classification procedure in-
a potential fall is detected. However, this approach focuses cludes five major steps: data preprocessing, image segmen-
only on detecting falls that occur while sitting down and tation, feature extraction, fall detection via GLR chart and
standing up from a chair. Others fall scenarios, such as falling classification by using the SVM algorithm (see Figure 1).
while picking something up from the ground and falling More details will be given in subsequent sections.
while getting out of bed, are not considered in this work. In
2 VOLUME 4, 2016
feature to detect a fall. However, the ellipse is not suitable for

detecting human movements because it cannot always dis-
criminate between different body postures (eg., dimensions,
and directionalities of the ellipse are too close for both sitting
FIGURE 1: Diagrammatic representation of camera-based and squatting positions, as illustrated in Figures 3 (b) and
fall detection technique. (c)). Accordingly, the idea of using a geometric shape has
been discarded in this study. In this work, we focused on non-
zero pixels constituting the human silhouette by the definition
A. PREPROCESSING AND SEGMENTATION of five occupancy regions. These regions were obtained by
The core role of segmentation, which is an important pre- a simple partitioning centered in the body’s center of mass.
liminary phase in computer vision applications, is to extract Partitioning is performed by setting five lines from the body’s
the monitored object from images. A silhouette is typically center of mass (see Figure 3 (d)). The first line is vertical,
discriminated by background subtraction [31], which uses two others segments are placed at 45◦ on either side of the
the background as a reference to removing invariable pixels vertical line and the two other segments are then put at 100◦
in the image sequence [31]. After segmentation, One can either side of the third and fourth segments, respectively.
usually observe some noise regions related to different as- The five areas in Figure 3 (d) A1 , A2 , A3 , A4 and A5 are
pects, such as shadows and lighting conditions. This noise colored cyan, green, yellow, blue and magenta respectively.
can be reduced or eliminated by applying morphological The center of mass and areas are calculated for each frame to
processing using erosion and dilation operators with 3 × 3 define features.
structuring elements. A sample of background subtraction
procedure is displayed in Figure 2. The background template
and the captured images are respectively given in Figure 2
(a) and (b). Results of the segmentation before and after
applying morphological operators are respectively illustrated
in Figure 2 (c) and (d).
FIGURE 3: Extracting features based on human body silhou-
ette.
Next, we compute normalized areas by dividing each

subarea value, Ai ; i = 1 . . . 5, by the total silhouette area,
A:
FIGURE 2: A sample of image segmentation technique. Ai
Ri = P5 . (1)
i=1 Ai
B. FEATURE EXTRACTION FROM IMAGES The extracted features are relevant and sufficient for de-
scribing the variation of human postures, in addition, they
Basically, feature extraction, which is an essential step in for
are easy to compute. The five ratios are computed for each
vision-based fall detection, consists in uncovering relevant
image and used as input for fall detection and classification
features from a segmented body. These attributes should be
(Figure 4).
invariant within the translation and scale changes. Recently,
numerous approaches for fall detection and classification
have been designed using shape information. In [32], [33],
the distance from the body’s gravity center to the floor is
presented as a fall indicator. In [25], Bian et al. proposed
a randomized decision tree algorithm for extracting the key
joints. However, the number of detected joints varies among
frames producing feature vectors of different sizes from one
frame to another. This approach would be inappropriate for
detection and classification techniques requiring fixed size of FIGURE 4: Features used to detect a fall.
the feature vector remain. In [34], Anderson et al. proposed
the bounding box as a descriptor of fall detection. In fact,
the horizontal and vertical dimensions of the bounding box 1) Fall detection indicator
are expected to change in the case of a fall event. Despite its Figure 5 displays an example of how the extracted features
simplicity, however, this approach is limited by its sensitivity (five areas) evolute when the inspected person performs
to carried objects, as shown in Figure 3(a). Motivated by typical movements including standing up, bending, picking
the limitations of the bounding box, in [35] Rougier et up an object, sitting down and lying down. Areas A3 and A5
al. proposed an approximated ellipse of the silhouette as a appear to be more sensitive to fall events than other areas
VOLUME 4, 2016 3
(Figure 5). This is observed by the magnitude of space in

these areas compared with the plot in other areas, indicating
their usefulness in fall detection.
FIGURE 5: Samples of human body partitioning into five

sections. Subject is standing up, picking up an object, sitting
down, and lying down.
Rather than using the vectors R3 and/or R5 for fall de-

tection, the 2-norm of R3 and R5 could be utilized as an
FIGURE 7: (a) Features extracted from data with fall for
indicator to detect falls. Let us define a fall indicator (FI)
frame ranging from 430 to 490. (b) The fall indicator under
as the square root of the summation of the squared of the
the presence of this failing situation.
normalized R3 (t) and R5 (t):
q
F I(t) = R32 (t) + R52 (t), t = 1, . . . , n, (2) In the next section, the GLR detector will be combined
with the SVM classifier for superior detection and classifica-
where n represents the number of samples or frames.
tion of falls.
Figures 6 (a) and (b) show the respective time evolution
of the five features (R1 − R5) and the fall indicator while III. ENHANCED MONITORING USING GLR ALGORITHM
performing commonplace activities. When the monitored AND SVM CLASIFICATION
person performs these daily activities (no falls), the fall This section presents firstly an overview of the GLR chart and
indicator defined by equation 2 is relatively small. However, how it can be employed in fall detection. Then, the proposed
when a fall occurs, the fall indicator increases significantly, fall detection approach, which integrates the GLR chart with
signalizing the occurrence of a new situation that is impor- SVM classifier is briefly presented.
tantly distinguishable from the activities of daily living (see
Figure 7 (b)). A. GLR-BASED DETECTOR
The GLR detector, an effective parametric hypothesis tests,
is employed for fall detection. It is extensively used in
several applications including air quality monitoring [16]
and train safety [28], [29]. The core role of the GLR ap-
proach is to discriminate between two composite hypotheses
H0 and H1 based on the observed data. For the purpose
of anomaly detection, let’s consider an observation vector
Y = [y1 , y2 , . . . , yn ] ∈ Rn being generated by one of these
Gaussian distributions:
H0 = {Y ∼ N (0, σ 2 In )}

(3)
H1 = {Y ∼ N (θ 6= 0, σ 2 In )},
where θ is the value of the anomaly and σ 2 > 0 is the
variance. In this paper, the null hypothesis, H0 , represents the
fall-free situation, and the alternative hypothesis, H1 , repre-
sents the situation with potential falls. Generally speaking,
to decide between the two hypotheses, the GLR approach
compares the decision statistic, L(Y ), to the control limit,
FIGURE 6: (a) Features extracted from data of activities h(α).
of daily living without fall (walking, standing, bending and ( sup fθ (Y )
siting, standing up). (b) The corresponding fall indicator, F I. H0 if L(Y ) = 2 log θ∈R
n
δ(Y ) = fθ=0 (Y ) < h(α) (4)
H1 else.
4 VOLUME 4, 2016
The GLR charting statistic, L(Y ), is given as Step 1 Collect the fall-free data (i.e., daily activities with-
out falls). The training data consists of the ex-
kY − θk22 kY k22

L(Y ) = 2 log sup exp − / exp − , tracted area ratios when the monitored person per-
2σ 2 2σ 2
θ forms daily activities. This is needed to determine
the thresholds.
where k . k2 is the Euclidean norm and fθ (Y ) =
1 Step 2 Scale the data to get variables with a mean of zero
n exp − 2σ1 2 k Y − θ k22 is the pdf of Y . Then, (5)
(2π) 2 σ n and variance of unity.
can be expressed as Step 3 Compute the fall indicator as given in equation (2)
1
using scaled fall-free data.
2 2
L(Y ) = min kY − θk 2 + kY k2 (5) Step 4 Compute the GLR threshold.
σ2 θ
1 n o 4) Testing phase:
= kY − b 2 + kY k2 .
θk 2 2 Step 1 Scale the testing data based on the mean and
σ2
variance of the training data.
After the estimation of θ as θb = arg minkY − θk22 = Y , Step 2 Compute the GLR statistic.
θ
L(Y ) can be expressed as Step 3 Flag a fall event when the GLR decision rule
overpass the threshold previously computed in the
1
L(Y ) = kY k22 . (6) training phase.
σ2
Regrettably, the actual drawback of this approach remains
The control limit, h(α), is defined to achieve the desired in its incapability to differentiate between true fall and false
probability of false alarms, selected a priori. falls (fall-like actions). To bypass such a limitation, fall
Z ∞ classification via the SVM algorithm will be used in this
P0 (L(Y ) ≥ h(α)) = f0 (y)dy=1 − Fχ21 (h) = α.(7) study.
h
C. SUPPORT-VECTOR-MACHINE-BASED FALL
Of course, an alarm is triggered by the GLR approach CLASSIFICATION
when the decision statistic, L(Y ), exceed the control limit, The SVM procedure was originally designed by Vapnik [36],
h(α). Otherwise, the supervised person is performing daily and has been extensively employed in classification [37],
activities. [38]. Basically, the key concept of the SVM is projecting
the input features in a high dimensional space via kernel
B. THE PROPOSED GLR-BASED FALL DETECTION functions and performing classification in the new space.
SCHEME This projection enables obtention of features relatively linear
The GLR procedure is applied to evaluate fall indicator. which facilitate the classification task. Then, the optimal
Using fall-free data, the GLR threshold is determined and hyperplane is determined in the transformed space to dis-
then used to evaluate the testing data. Falls are flagged when criminate input observations [36]. Here, the SVM procedure
the GLR statistic surpasses the decision threshold. Figure 8 is used for fall classification due to its flexibility to deal
presents a block diagram of the proposed procedure and the with linear and nonlinear data by using nonlinear kernels.
main steps to implement this approach are outlined next. There are numerous kernel functions that can be used in the
classification step. In this study, we tested three commonly
used kernels: linear, polynomial, and radial basis function.
To evaluate the capability of the proposed technique, we
used 3-fold cross-validation in classification step and we
computed different statistical metrics to quantify the accu-
racy of our results, including the overall accuracy and the
FIGURE 8: GLR monitoring chart. Area Under Curve (AUC) [39]. Figure 9 illustrates a con-
fusion matrix and summarizes equations of the main related
1) Segmentation: A background subtraction procedure is metrics that are commonly used to assess the quality of a
used to extract the body’s silhouette from the image binary decision method and which will be used to assess the
sequence, and morphological processing is used for performance of the GLR-SVM based fall detection approach.
removing or reducing noise.
2) Human body feature extraction: The five areas of the In the next section, the performance of the GLR-SVM
monitored human body are extracted. For each frame, approach will be assessed and compared to that of the
we compute a set of ratios and concatenate them to state-of-the-art fall detection techniques through two publicly
form the feature vector. The fall indicator used in fall available datasets.
detection, F I, comprises the 2-norm of A3 and A5 .
3) Training phase: IV. EXPERIMENTS
VOLUME 4, 2016 5
FIGURE 9: Confusion matrix and associated commonly used performance metrics.
A. DATA DESCRIPTION
To assess the ability of the GLR scheme to identify a fall,
experiments were conducted on two known databases: the
University of Rzeszow’s fall detection data (URFD) and
the fall detection data (FDD) [26], [40]. URFD consists
of 70 video sequences including several actions performed
in different ways. In this work, all fall incidents and daily
activities were acquired from the RGB camera. The second
dataset FDD contains 191 videos with a rate of 25 images /s
and a resolution of 320×240 pixels.
The GLR scheme was performed on the training fall-
free data. A 1024 fall-free sub-videos have been selected
for training. The extracted features from the training data
organized as a matrix of 1024 rows (frames) and 5 columns
(number of features). These data were scaled and then used
to compute the fall indicator, which is defined by equation 2. FIGURE 10: Top: The features of testing data with fall for
The GLR threshold is set up to be h = 1.79 to achieve a frame ranging from 430 to the final frame. Bottom: GLR
probability of false alarm of 5%. statistic for testing data.
B. DETECTION RESULTS
Here, two cases of falls have been investigated. The first one
contains a real fall in the testing sequence (case A), and the
second one includes a false fall (case B).
1) Scenarios with true fall -case A

Three scenarios are presented here to test the feasibility of
the GLR procedure in identifying falls. In the first scenario,
a fall happened in the testing data (comprising 490 frames)
for frame ranging from 430 to 490 (see the top panel of
Figure 10). Note that the GLR procedure is applied to the
fall indicator (FI) obtained using equation 2. The results of
the GLR procedure are illustrated in the bottom panel of
Figure 10 and show its capability to detect accurately the fall.
In the second scenario, the test data comprise a true fall FIGURE 11: Top: the testing data contain true fall for frame
between frame 46 and the final frame (the top panel of ranging from 46 to 170. Bottom: the time evolution of the
Figure 11). The GLR chart (the bottom panel of Figure 11) GLR statistic when it is applied to FI obtained from the data
successfully detected this fall. given in top of this figure.
In the third scenario, the test data contains a fall event for
frame ranging from 150 to 385 (see the top of Figure 13).
Results of the GLR chart, displayed in the bottom panel of the top panel of Figure 13). The GLR chart depicted in the
Figure 13, indicate that the fall was correctly detected. bottom panel of Figure 13 flagged this as a fall.
Therefore the GLR chart cannot identify the type of this
2) Scenarios with false falls -Case B abnormal event. To deal with this difficulty, a fall detection
In this scenario, the test measurements contain false fall task is then followed by a classification phase.
(lying down) events for frame ranging from 180 to 385 (see
6 VOLUME 4, 2016
In practice, the kernel function has an important role in

the performance of the SVM classifier. As the classification
performance of the SVM algorithm depends on the selected
kernel function, in this study, SVM with three frequently used
kernel functions (linear, RBF, and polynomial kernels) are
compared.
- Linear kernel:
K(xi , xj ) = xi .xj ,
- Polynomial kernel:
K(xi , xj ) = (xi .xj + 1)d ,
where d is the degree of polynomial kernel.
- Radial basis function kernel:
FIGURE 12: Top: The testing data contain a true fall between
frame numbers 150 and 385. Bottom: The time evolution kxi − xj k2
K(xi , xj )=exp( ),
of the GLR statistic when it is applied to the fall indicator 2σ 2
obtained from the data given in the top of this figure. where σ is the width of Gaussian kernel.
An example of SVM classification using the three kernel
functions is given in Figure 14. To reduce the dimensions of
visualization, we used principal component analysis and plot
the first and second principal components (Figure 14).
FIGURE 13: Top: The testing data contain a true fall for
frame ranging from 180 to 385. Bottom: The time evolution
of the GLR statistic when it is applied to the fall indicator
obtained from the data given at the top of this figure.
FIGURE 14: Classification results of SVM using each kernel

C. CLASSIFICATION RESULTS function: linear (a), polynomial (b) and RBF (c).
As mentioned above, during the GLR detection step, a con-
fusion between real falls and certain like-falls activities are After testing the three kernel functions with SVM, the best
observed. To mitigate this problem, the SVM classification accuracy is realized when using the RBF kernel function,
phase is then launched when a potential fall is flagged by which provided a classification accuracy of 96.66%(Table 1).
the GLR approach. In other words, the SVM algorithm is not Figure 15 displays ROC curves corresponding to SVM with
trained on sequences corresponding to daily activities, but it linear, polynomial and RBF kernel functions when features
focuses only on detected falls (i.e., sequences corresponding of detected falls by GLR chart have been used as inputs to the
to false and real falls). In the SVM-based classification, SVM classifier. One can see that the SVM classification using
all five features are used to discriminate between true and RBF kernel function performed better than the other kernel
like fall cases. The SVM classifier’s input is constituted by functions(AUClienar = 0.9405; AUCpolynomial = 0.9435
the concatenation of extracted features from each frame. To AUCRBF = 0.9526). Also, we notice from the shape of
verify the efficiency of the SVM algorithm, classification is the ROC plots, higher sensitivities with RBF kernel function
performed using a three-fold cross-validation technique. were obtained at higher specificities compared to the other
VOLUME 4, 2016 7
two kernels (Table 1). In this case, the RBF kernel function
proved to be the best reliable kernel function. Note that the
data can be seen as non-linearly separable, making a linear
kernel function unsuitable.
TABLE 2: Performance comparison between the studied
algorithms.
of the daily activities from falling cases and thus reduc-

ing the amount of data to be utilized in the classifica-
tion stage. As shown in TAbletableResults, the SVM-GLR
achieves the highest AUC of 0.97 and outperforms the other
studied classifiers (AUCKN N = 0.93; AUCN N = 0.94
AUCN aive Bayes = 0.95, and AUCDecision T ree = 0.92).
In addition, SVM formalism is less prone to overfitting.
GLR-SVM outperformed k-NN due to SVMs’ capacities for
representing data in comparison to the complexity of k-NN in
finding the k nearest neighbors for each new video sequence.
FIGURE 15: ROC curves for SVM with linear, third-order The GLR-SVM provides efficient detection results compared
polynomial, and RBF (σ = 0.125) kernel functions, using all to the decision trees algorithm because of its sensibility to
features corresponding to the detected fall. small variations in the features. Then, the classification result
can be completely different when the feature set is treated
by other generated trees. The GLR-SVM has also surpassed
TABLE 1: Performance comparison of SVM algorithms us- Naïve Bayes algorithm. This is mainly related to the indepen-
ing different kernel functions: linear, polynomial and RBF. dence assumption which is far getting and usually inappro-
priate with real applications. Furthermore, SVMs are applied
only on the detected sequences, so just a reduced number of
video sequences is treated in the classification step, which
makes the processing simple. Instead of the fall classification
approaches, where all video sequences are concerned by the
Table 2 illustrates the comparison between the integrated classification (a larger number of scenarios of activities are
GLR-SVM method with some frequently used standalone trained), this complex learning makes classification without
classifiers, k-nearest neighbor, neural network, naïve Bayes detection phase inappropriate for the application at hand.
and decision tree, without detection step. It is noteworthy Moreover, Tables 3 and 4 show that better performances
that the classifier’s parameters have an important influence are obtained with the GLR-SVM compared to some state-
on the classification performance. During the learning phase, of-the-art machine learning procedures when using FDD and
for each classifier, the optimal parameters providing the URFD databases. We also notice that the integrated GLR-
highest accuracies are selected. For the multilayer perceptron SVM mechanism is much reliable in detecting falls than
neural network, the optimal parameters which correspond to applying a standalone classifier. Merging the detection step
the highest accuracy contain 3 layers, including 15 neurons with the classification step permits to separate daily activities
(nodes) in the hidden layer have been selected. The number from falls, and then reduce the input features used by the
of neighbors, k, in k-NN has been varied from 1 to 20 to SVM-based classification. In the SVM algorithm, the data is
select the best value. The value k = 3 provides the highest transformed into a higher dimensional space using a kernel
accuracy. In the case of the decision tree, the regression is function to appropriately separate falls and like-falls data.
selected as a classification method and a maximum deviance Additionally, SVM classification provides a sparse solution
reduction as a criterion for choosing a split. Finally, for SVM via structural risk minimization, unlike neural networks,
classification, the kernel’s parameters (the cost C and the which is based on empirical risk minimization.
width of Gaussian kernel σ) are first varied and the cou-
ple corresponding to the best performance has been chosen V. CONCLUSION
(σ = 0.125 and C = 128). In this study, a statistical strategy for efficient fall detection
Results in Table 2 confirm also the feasibility and su- and classification using a GLR chart and SVM classifier is
periority of the integrated GLR-SVM method compared to presented. First, the human body is divided into five portions
the standalone classifiers used in this study. The GLR-SVM corresponding to five partial occupancy areas. Then, for
demonstrated good capacity in discriminating real fall from each frame, the area ratios (extracted features) have been
fall-like actions. This highlights the benefit of the GLR- computed and utilized as input data. The feasibility of the
based fall detection step, which consists in separating most designed mechanism was tested on two publicly available
8 VOLUME 4, 2016
TABLE 3: Performance comparison based on FDD database.

Approach Detector Overall accuracy(%)
L. Alhimale et. Al [41] Bounding box features NN 94.27%
X. Ma el al. [42] Silhouette, lighting and flow features VPSO-NN 92.17%
M. Yu et al. [43] Projection, histogram along the axes of the ellipse DAGSVM 96.09%
GLR-SVM Fall index and area ratios GLR-SVM 96.84%
TABLE 4: Performance comparison based on URFD database.
datasets. The results demonstrated that the combined GLR [10] S. Zhang, P. J. McCullagh, H. Zheng, and C. Nugent, “Situation aware-
detection scheme with the SVM classification was not only ness inferred from posture transition and location; derived from smart
phone and smart home sensors,” IEEE Transactions on Human-Machine
able to separate between daily activities and falls but it was Systems, vol. 47, no. 6, pp. 814–821, 2017.
also able to distinguish true false from false fall events (like [11] L. Montanini, A. Del Campo, D. Perla, S. Spinsante, and E. Gambi, “A
intentional lying or missing step). Data were also collected to footwear-based methodology for fall detection,” IEEE Sensors Journal,
vol. 18, no. 3, pp. 1233–1242, 2018.
evaluate the combined GLR-SVM fall detection strategy in
[12] B. Kwolek and M. Kepski, “Improving fall detection by the use of depth
comparison to other classifiers including neural network, K- sensor and accelerometer,” Neurocomputing, vol. 168, pp. 637–645, 2015.
nearest neighbor, decision tree, and naïve Bayes classifiers. [13] W.-Y. Shieh and J.-C. Huang, “Falling-incident detection and throughput
The major finding was that the combination of GLR and enhancement in a multi-camera video-surveillance system,” Medical engi-
neering & physics, vol. 34, no. 7, pp. 954–963, 2012.
SVM techniques provided a method with fewer false-positive [14] S.-G. Miaou, P.-H. Sung, and C.-Y. Huang, “A customized human fall
falls than other approaches, making it a more accurate fall detection system using omni-camera images and personal information,”
detection technique. in Distributed Diagnosis and Home Healthcare, 2006. D2H2. 1st Transdis-
ciplinary Conference on. IEEE, 2006, pp. 39–42.
As future work, we plan to incorporate information from [15] B. Jansen and R. Deklerck, “Context aware inactivity recognition for
Kinect camera (instead of RGB camera), which operates visual fall detection,” in Pervasive Health Conference and Workshops,
2006. IEEE, 2006, pp. 1–4.
on depth imagery and preserves privacy for people being
[16] F. Harrou, L. Fillatre, M. Bobbia, and I. Nikiforov, “Statistical detection
monitored. Hence, in this case, the proposed fall detection of abnormal ozone measurements based on constrained generalized likeli-
strategy can be used in a similar manner as in the case of hood ratio test,” in Decision and Control (CDC), 2013 IEEE 52nd Annual
RGB cameras. Conference on. IEEE, 2013, pp. 4997–5002.
[17] P. Dašić, J. Dašić, and B. Crvenković, “Improving patient safety in hos-
pitals through usage of cloud supported video surveillance,” Open access
REFERENCES Macedonian journal of medical sciences, vol. 5, no. 2, p. 101, 2017.
[1] W. H. Organization, W. H. O. Ageing, and L. C. Unit, WHO global report [18] G. A. W. Index, “Global age watch index 2015: Insight report,” London:
on falls prevention in older age. World Health Organization, 2008. HelpAge International, 2015.
[2] L. Ren and Y. Peng, “Research of fall detection and fall prevention [19] C. Barelle, E. Vellidou, and D. Koutsouris, “The BIOTELEKINESY home
technologies: A systematic review,” IEEE Access, vol. 7, pp. 77 702– care service: A holistic concept to prevent fall among the elderly based on
77 722, 2019. ICT and computer vision,” in Proceedings of the 10th IEEE International
[3] C. Taramasco, T. Rodenas, F. Martinez, P. Fuentes, R. Munoz, R. Olivares, Conference on Information Technology and Applications in Biomedicine.
V. H. C. De Albuquerque, and J. Demongeot, “A novel monitoring system IEEE, 2010, pp. 1–5.
for fall detection in older people,” IEEE Access, vol. 6, pp. 43 563–43 574, [20] K. Ozcan, S. Velipasalar, and P. K. Varshney, “Autonomous fall detection
2018. with wearable cameras by using relative entropy distance measure.” IEEE
[4] A. Lotfi, S. Albawendi, H. Powell, K. Appiah, and C. Langensiepen, Trans. Human-Machine Systems, vol. 47, no. 1, pp. 31–39, 2017.
“Supporting independent living for older adults; employing a visual based
[21] C. Rougier, J. Meunier, A. St-Arnaud, and J. Rousseau, “Fall detection
fall detection through analysing the motion and shape of the human body,”
from human shape and motion history using video surveillance,” in
IEEE Access, vol. 6, pp. 70 272–70 282, 2018.
Advanced Information Networking and Applications Workshops, 2007,
[5] E. Cippitelli, F. Fioranelli, E. Gambi, and S. Spinsante, “Radar and rgb-
AINAW’07. 21st International Conference on, vol. 2. IEEE, 2007, pp.
depth sensors for fall detection: a review,” IEEE Sensors Journal, vol. 17,
875–880.
no. 12, pp. 3585–3604, 2017.
[6] A. Chelli and M. Pätzold, “A machine learning approach for fall detection [22] S.-G. Miaou, F.-C. Shih, and C.-Y. Huang, “A smart vision-based human
and daily living activity recognition,” IEEE Access, vol. 7, pp. 38 670– fall detection system for telehealth applications,” in Proc. Third ISATED
38 687, 2019. Int. Conf. on Telehealth, 2007, pp. 7–12.
[7] P. Pierleoni, A. Belli, L. Maurizi, L. Palma, L. Pernini, M. Paniccia, and [23] E. Auvinet, F. Multon, A. Saint-Arnaud, J. Rousseau, and J. Meunier, “Fall
S. Valenti, “A wearable fall detector for elderly people based on ahrs and detection with multiple cameras: An occlusion-resistant method based on
barometric sensor,” IEEE Sensors Journal, vol. 16, no. 17, pp. 6733–6744, 3-D silhouette vertical distribution,” IEEE Transactions on Information
2016. Technology in Biomedicine, vol. 15, no. 2, pp. 290–300, 2011.
[8] Y.-Z. Hsieh and Y.-L. Jeng, “Development of home intelligent fall de- [24] H. Foroughi, B. S. Aski, and H. Pourreza, “Intelligent video surveillance
tection iot system based on feedback optical flow convolutional neural for monitoring fall detection of elderly in home environments,” in Com-
network,” IEEE Access, vol. 6, pp. 6048–6057, 2017. puter and Information Technology, 2008. ICCIT 2008. 11th International
[9] Y. Zigel, D. Litvak, and I. Gannot, “A method for automatic fall detection Conference on. IEEE, 2008, pp. 219–224.
of elderly people using floor vibrations and sound-proof of concept on hu- [25] Z.-P. Bian, J. Hou, L.-P. Chau, and N. Magnenat-Thalmann, “Fall detection
man mimicking doll falls,” IEEE Transactions on Biomedical Engineering, based on body part tracking using a depth camera,” IEEE journal of
vol. 56, no. 12, pp. 2858–2867, 2009. biomedical and health informatics, vol. 19, no. 2, pp. 430–439, 2015.
VOLUME 4, 2016 9
[26] B. Kwolek and M. Kepski, “Human fall detection on embedded platform

using depth maps and wireless accelerometer,” Computer methods and
programs in biomedicine, vol. 117, no. 3, pp. 489–501, 2014.
[27] K.-L. Lu and E. Chu, “An image-based fall detection system for the
elderly,” Applied Sciences, vol. 8, no. 10, p. 1995, 2018.
[28] F. Harrou, L. Fillatre, and I. Nikiforov, “Anomaly detection/detectability
for a linear model with a bounded nuisance parameter,” Annual Reviews
in Control, vol. 38, no. 1, pp. 32–44, 2014.
[29] ——, “Anomaly detection with bounded nuisance parameters and safe
train navigation,” in 2008 10th International Conference on Control,
Automation, Robotics and Vision. IEEE, 2008, pp. 1021–1026.
[30] S. Wang, L. Chen, Z. Zhou, X. Sun, and J. Dong, “Human fall detection in
surveillance video based on PCANet,” Multimedia Tools and Applications,
pp. 1–11, 2015.
[31] K. Kim, T. Chalidabhongse, D. Harwood, and L. Davis, “Real-time
foreground–background segmentation using codebook model,” Real-time
imaging, vol. 11, no. 3, pp. 172–185, 2005.
[32] B. Kwolek and M. Kepski, “Fuzzy inference-based fall detection using
kinect and body-worn accelerometer,” Applied Soft Computing, vol. 40,
pp. 305–318, 2016.
[33] H. Liu and C. Zuo, “An improved algorithm of automatic fall detection,”
AASRI Procedia, vol. 1, pp. 353–358, 2012.
[34] D. Anderson, J. Keller, M. Skubic, X. Chen, and Z. He, “Recognizing falls
from silhouettes,” in 28th Annual International Conference of the IEEE
Engineering in Medicine and Biology Society, EMBS’06. IEEE, 2006,
pp. 6388–6391.
[35] C. Rougier, J. Meunier, A. St-Arnaud, and J. Rousseau, “Robust video
surveillance for fall detection based on human shape deformation,” IEEE
Transactions on Circuits and Systems for Video Technology, vol. 21, no. 5,
pp. 611–622, 2011.
[36] V. Vapnik, The nature of statistical learning theory. Springer Science &
Business Media, 2013.
[37] X. Gao and J. Hou, “An improved SVM integrated GS-PCA fault diagnosis
approach of tennessee eastman process,” Neurocomputing, vol. 174, pp.
906–911, 2016.
[38] Z. Yin and J. Hou, “Recent advances on SVM based fault diagnosis and
process monitoring in complicated industrial processes,” Neurocomputing,
vol. 174, pp. 643–650, 2016.
[39] N. Zerrouki and A. Houacine, “Automatic classification of human body
postures based on curvelet transform,” in Image Analysis and Recognition.
Springer, 2014, pp. 329–337.
[40] I. Charfi, J. Miteran, J. Dubois, M. Atri, and R. Tourki, “Definition and
performance evaluation of a robust SVM based fall detection solution,”
in Signal Image Technology and Internet Based Systems (SITIS), 2012
Eighth International Conference on. IEEE, 2012, pp. 218–224.
[41] L. Alhimale, H. Zedan, and A. Al-Bayatti, “The implementation of an
intelligent and video-based fall detection system using a neural network,”
Applied Soft Computing, vol. 18, pp. 59–69, 2014.
[42] X. Ma, H. Wang, B. Xue, M. Zhou, B. Ji, and Y. Li, “Depth-based human
fall detection via shape features and improved extreme learning machine,”
IEEE journal of biomedical and health informatics, vol. 18, no. 6, pp.
1915–1922, 2014.
[43] M. Yu, A. Rhuma, S. M. Naqvi, L. Wang, and J. Chambers, “A posture
recognition-based fall detection system for monitoring an elderly person in
a smart home environment,” IEEE transactions on information technology
in biomedicine, vol. 16, no. 6, pp. 1274–1286, 2012.
10 VOLUME 4, 2016

An Integrated Vision-Based Approach For Efficient

Uploaded by

Copyright:

Available Formats

An Integrated Vision-Based Approach For Efficient

Uploaded by

Document Information

Original Title

Copyright

Available Formats

Share this document

Share or Embed Document

Sharing Options

Did you find this document useful?

Is this content inappropriate?

Copyright:

Available Formats

An Integrated Vision-Based Approach For Efficient

Uploaded by

Copyright:

Available Formats

This article has been accepted for publication in a future issue of this journal, but has not been

An integrated vision-based approach for

feature to detect a fall. However, the ellipse is not suitable for

Next, we compute normalized areas by dividing each

(Figure 5). This is observed by the magnitude of space in

FIGURE 5: Samples of human body partitioning into five

Rather than using the vectors R3 and/or R5 for fall de-

FIGURE 9: Confusion matrix and associated commonly used performance metrics.

1) Scenarios with true fall -case A

In practice, the kernel function has an important role in

FIGURE 14: Classification results of SVM using each kernel

of the daily activities from falling cases and thus reduc-

TABLE 3: Performance comparison based on FDD database.

TABLE 4: Performance comparison based on URFD database.

[26] B. Kwolek and M. Kepski, “Human fall detection on embedded platform

You might also like