A fall detection system using k-nearest neighbor classifier

(1)

A fall detection system using k-nearest neighbor classiﬁer

Chien-Liang Liu

*

_{, Chia-Hoang Lee, Ping-Min Lin}

Department of Computer Science, National Chiao Tung University, 1001 University Road, Hsinchu 300, Taiwan, ROC

a r t i c l e

i n f o

Keywords: Artiﬁcial intelligence Machine learning kNN Fall detection

a b s t r a c t

The main purpose of this paper is to use off-the-shelf devices to develop a fall detection system. In human body identification, human body silhouette is adopted to improve privacy protection, and vertical projec-tion histograms of the silhouette image and statistical scheme are used to reduce the effect of human body upper limb activities. The kNN classification algorithm is used to classify the postures using the ratio and difference of human body silhouette bounding box height and width. Meanwhile, since time differ-ence is a vital factor to differentiate fall incident event and lying down event, the critical time differdiffer-ence is obtained from the experiment and verified by statistical hypothesis testing. With the help of the kNN classifier and the critical time difference, a fall incident detection system is developed to detect fall inci-dent events. The experiment shows that it could reduce the effect of upper limb activities and the system has a correct rate of 84.44% on fall detection and lying down event detection.

1. Introduction

During the 20th century, the US population under age 65 tri-pled, but those who were 65 and older increased by a factor of 11. It is apparent that the proportion of older people is rising. About 1 in 8 Americans were elderly in 1994, but about 1 in 5 would be elderly by the year 2030 (Bureau, 2008). As with the great progress of medical technology, many countries are facing the issue of aging society, so there will be a lower proportion of people of working age available both to fund and to provide the necessary levels of care. Meanwhile, the problem comes after this issue is huge nursing cost and both of them inﬂuence our life now. Therefore, the demand for developing home surveillance systems is rising.

It is shown in National Center for Health Statistics (2000) and DA and JA (2001) that more than one-third of adults ages 65 and older fall each year. Falls are the leading cause of fatal and non-fatal injuries for people aged 65 and older. For adults 65 years old or older, 60% of fatal falls happen at home, 30% occur in public places, and 10% occur in health care institutions (GS, 1988). Re-cently, a lot of fall detection systems and posture classiﬁcation sys-tems have been developed to detect fall incidents (Anderson, Keller, Skubic, Chen, & He, 2006;Bramberger, Doblander, Maier, Rinner, & Schwabach, 2006; Cucchiara, Grana, Prati, & Vezzani, 2005; Juang & Chang, 2007; Miaou, Sung, & Huang, 2006; Nasution & Emmanuel, 2007; Tabar, Keshavarz, & Aghajan, 2006; Tao, Turjo, Wong, Wang, & Tan, 2005). Instead of using specially designed sensors and circuitry (Noury et al., 2000; Tamura, Yoshimura,

Horiuchi, Higashi, & Fujimoto, 2000), the purpose of this study is to focus on using off-the-shelf devices to develop a surveillance system to detect fall incidents. The motives of our work arises from the fact that current home surveillance system is quite expensive and most people could not afford to buy one. Meanwhile, computer and web camera equipments are popularized recently and it will be of great help if these equipments could be used to provide sur-veillance service.

In essence, fall detection includes two main tasks: human body identification and fall incident detection. Identifying moving ob-jects from a video sequence is a fundamental and critical task in many computer-vision applications such as video surveillance and traffic monitoring systems. Practically, video surveillance sys-tem may cause privacy threat to those who use the syssys-tem, so hu-man body silhouette is adopted here to improve privacy protection. Since human body silhouette has removed most of hu-man body features, huhu-man body identification is prone to be ef-fected by human actions. For example, people in the indoor may stretch out their hands to fetch and that may affect human body posture identification. Pixels projection and statistical analysis are used to reduce the effect.

The k-Nearest Neighbor (kNN) classiﬁer (Duda, Hart, & Stork, 2000) is adopted to classify user’s activities based on the features. In kNN classiﬁcation algorithm, the number of classes and feature selection will be key factors. k-fold cross-validation is adopted to choose the best k value (the number of classes), and statistical analysis is used to facilitate feature selection. Practically, lying down event and falling down event are so similar that it may often lead to misjudgment. According to our experiments, the time dif-ference between time of lying posture and temporary posture could be used to differentiate these two activities. In this paper,

*Corresponding author. Tel.: +886 3 5131503; fax: +886 3 5734935. E-mail address:[email protected](C.-L. Liu).

Contents lists available atScienceDirect

Expert Systems with Applications

(2)

the critical value of time difference is determined by experiments and veriﬁed by statistical hypothesis testing. According to the experiment, the system could successfully detect fall incident and lying down event with a correct rate of 84.44%.

The rest of the paper is structured as follows. In Section2a sur-vey of related researches on indoor monitoring and surveillance is presented. Section3 describes the system framework which in-cludes the human body identiﬁcation and fall incident detection. In Section4, several experiments are conducted to show that it works well on fall incident detection. Finally, Section5 contains the conclusion.

2. Preliminary

Rapid advances in the technologies of image sensors and embed-ded processors enable the inclusion of vision-based nodes in vari-ous smart environment applications (Bramberger et al., 2006). Various studies on the use of accelerometer signal for posture clas-siﬁcation have been reported.Tabar et al. (2006)proposed a wire-less sensor network employing multiple sensing and event detection modalities for smart home monitoring applications. The fall alert is broadcasted by the user badge which is carried by the user when the accelerometers on the badge record a signiﬁcant change in their measured signal. Meanwhile, image processing is used to analyze the situation and determine the user’s posture when alerts happen.Miaou et al. (2006)proposed to use MapCam (omni-camera) to detect the fall of the elderly. Moreover, the per-sonal information of each individual is considered in the processing task and the MapCam could capture 360° sense simultaneously.

Tao et al. (2005)proposed an intelligent video surveillance sys-tem to detect human fall incidents for enhanced safety in indoor environments and the feature they use is the aspect ratio of the moving object’s bounding box. The system proposed byTao et al. (2005)consists of two main parts: a vision component which can reliably detect and track moving people in the view of a camera, and an event-inference module which parses observation se-quences of people features for possible falling behavioral signs.

Anderson et al. (2006)adopted the width to height ratio of the silhouette bounding box and the off-diagonal term from the covari-ance matrix as the features to determine whether fall incident oc-curs. The bounding box width to height ratio indicates whether the silhouette is larger in the vertical plane versus the horizontal plane. When the subject is standing, the width to height ratio is small. Meanwhile, when the subject is on the ground, the width to height ratio becomes much larger. These features need to be extracted from the silhouette to train and perform classiﬁcation with HMMs for temporal pattern recognition.

In practice, privacy should be protected when designing fall detection systems. Currently many fall detection systems adopt sil-houette to remove most of the human body features and human body identiﬁcation will be prone to be effected by human actions. Therefore, it is necessary to reduce the effects of such actions. Be-sides, lying down and fall incidents events are similar and the sys-tem should be able to differentiate these two activities to avoid false alarms. These existing fall detection schemes (Miaou et al., 2006; Nait-Charif & McKenna, 2004; Tao et al., 2005) are unable to differentiate between real fall incident and an event where the person is simply lying without really falling (Nasution & Emmanu-el, 2007).

3. System architecture

The falling detection system in this paper consists of two main parts: human body identiﬁcation and fall incident detection. The objective of the human body identiﬁcation is to identify human

body and extract useful features from human body images. Mean-while, fall incident detection will make use of these features to determine whether fall incident is taking place. Fig. 1shows the system architecture in this paper. The input is the video image frames that are obtained from web camera. Video surveillance sys-tem may infringe privacy to those who use the syssys-tem, so human body silhouette is adopted to improve privacy protection and sev-eral image-processing techniques are used to identify the object and to smooth noises. Feature selection is commonly used in ma-chine learning applications for building robust learning models. Silhouette of the human body removes most of the human body features, so upper limb activities may affect the identification re-sult and it is necessary to reduce the effects of such actions. The kNN classification scheme is used to perform postures classifica-tion, and 450 sample data is used as training data to build the training model. Finally, fall incident detection is achieved by fall incident detection flow.

3.1. Human body identiﬁcation

As mentioned above, moving objects should be identiﬁed from the videos for further analysis. Frame differencing approach, which is low-complexity, is adopted to identify human body. Fig. 2(a) shows the image captured from the camera andFig. 2(b) shows the human body identiﬁcation result.

3.2. Image processing

The human body image extracted from moving object may con-tain noises, so additional process is required to eliminate the noises. Mean ﬁlter, which could soften an image by averaging sur-rounding pixel values, is adopted to make the image more smooth.

(3)

As shown inFig. 3, each pixel surrounding the target is assigned the same weight and the average of the pixels is assigned to the target.

The gray-scale image will be transformed into binary image for further process, and thresholding is a method which could trans-form a gray-scale image into binary image (Shapiro & Stockman,

2001). Binary images are images whose pixels have only two pos-sible intensity values. They are normally displayed as black and white. Numerically, the two values are often 0 for black, and either 1 or 255 for white.

In addition to frame differencing, mean ﬁlter, and image thres-holding processes, connected component labeling, which scans an image and group its pixels into components based on pixel connec-tivity, is used to capture the meaningful region in this paper. An arbitrary point on the image is marked as connecting objects and an unique number is assigned to all the connecting pixels that form four-neighbors to represent that they connect to each other. Once an unmarked pixel is discovered, a new number will be assigned to the pixel. The above process will repeat until all the pixels of the image have been marked. For each non-zero pixel, its four-neigh-bors will be checked, and if non-zero pixels are found, they will be at the same group and one number will be assigned as shown inFig. 4.

3.3. Features selection

After the above image-processing steps are performed, the im-age of human body silhouettes would be obtained. In essence, although human body silhouette could protect user’s privacy, hu-man body silhouette also removes most of huhu-man body features. As shown inFig. 5, the human body silhouette could not reveal per-sonal information except the height and width of the silhouette.

In general, indoor activities include walking, sitting, standing, and lying etc. In this paper, the indoor activities are divided into three categories: standing posture, temporary posture and lying down posture. As shown inFig. 6, aspect ratios of moving object’s bounding box in standing posture and lying down posture repre-sent two extreme values in indoor activities, while the temporary posture represents all other postures whose aspect ratios of mov-ing object’s boundmov-ing box are located between these two values. For example, temporary posture includes sitting, bending, and squatting postures etc.

Nasution and Emmanuel (2007) proposed to use horizontal and vertical projection histograms of segmented foreground and angle between last standing posture with current foreground bounding box as feature set for the task. Since projection hisgram of lying toward the camera is closer to bending, lying to-ward the camera may be detected as bending. Nasution and Emmanuel (2007)proposed a modiﬁed classiﬁer to improve the

Fig. 2. Video frame extraction from video and identiﬁcation of the human body.

Fig. 3. Mean ﬁlter.

Fig. 4. Connected components labeling using four-connected neighbor.

Fig. 5. Human body bounding box and its silhouette.

(4)

detection of lying toward the camera, but that will need to store and update the bounding box information of the last detected standing posture all the time. Instead of using horizontal and ver-tical projection histograms and angle as feature set, the ratio of bounding box height and width and the difference of bounding box height and width are adopted as the system features in this paper.Fig. 7shows the feature space result of experiment when adopting these new features.

3.4. Reduction of upper limb activities effect

As mentioned above, human body silhouette removed most of human body features, so height and width of human body silhou-ette are adopted as the features of fall incident detection in many studies. However, people may stretch out their hand to fetch some-thing in the indoor environment and these activities may lead to incorrect measurement of the width of human body silhouette im-age. Therefore, it is necessary to reduce the inﬂuence of upper limb activities to obtain an approximate width value which is close to the original one. In this paper, we propose to use vertical projection

histograms of the silhouette image and adopt statistical analysis to reduce the inﬂuence.Fig. 8shows the projection result.

As shown inFig. 9(a), the bounding box of the human body will change when the person stretch his hand and the vertical projec-tion histograms of the silhouette image will produce a peak in the distribution curve. The individual mean and standard deviation are calculated, and a new value will be used to replace the original extreme value which is over threshold value and the threshold va-lue is determined by Eq.(1). This procedure performs well to elim-inate the effect of upper limb activities.Fig. 9(b) shows the result of eliminating upper limb activities after using this approach. The peak section will be reduced when compared with the original curve

threshold ¼ 3

r

þ

l

: ð1Þ

3.5. kNN classiﬁer for human body postures classiﬁcation

The k-nearest neighbor (kNN) classiﬁer (Duda et al., 2000), a supervised machine learning technique for learning a function

Fig. 7. Feature space of sample data.

Fig. 8. Project bounding box image pixels onto Y-axis.

(5)

from training data, classiﬁes objects based on closest training data in the feature space. In order to identify neighbors, the objects could be represented by position vectors in a multidimensional feature space, and different distance measures have been proposed in the literature. Euclidean distance is used in this paper to repre-sent the distance between the objects.

As described above, the system in this paper divides the indoor activities into three categories: standing posture, temporary pos-ture, and lying pospos-ture, respectively. In this paper, 450 human body postures are collected for experiment, and each category con-tains 150 samples. In essence, a point in the space is assigned to the class c if it is the most frequent class label among the k nearest training samples, so the parameter k plays an important role in kNN classiﬁcation.

In essence, kNN classifier should determine the parameter k in advance and the best choice of k depends upon the data. In general, larger values of k will reduce the effect of noise on the classifica-tion, but cause boundaries between classes less distinct. In this pa-per, cross-validation is adopted to find out the best one for the classification. In k-fold cross-validation, the original sample is par-titioned into k subsamples. Of the k subsamples, a single subsam-ple is retained as the validation data for testing the model, and the remaining k 1 subsamples are used as training data. The cross-validation process is then repeated k times, with each of the k

subs-amples used exactly once as the validation data. The k results from the folds then can be averaged to produce a single estimation.

Table 1shows the result when adopting k-fold cross-validation to verify the correctness of our kNN classiﬁer with the new fea-tures. As shown inTable 1, the system could produce better result when k equals to 3, so k’s value could be determined as well.

3.6. Fall incident detection ﬂow

The fall detection system in this paper will focus on indoor activities and Fig. 10 describes the fall incident detection flow and the intermediate events include standing posture event, lying posture event, and temporary posture event, while the final events include falling down event and lying down event. The kNN classi-fier could identify the intermediate events and the decision of a fall incident is determined by the event transition and time difference between events. In essence, it may lead to false alarms on fall inci-dent event and lying down event, because these two events are similar. According to our experiments, the speed of transition is an important factor which can successfully distinguish these two events. In other words, if the transition time between temporary posture event and lying down event is less than certain threshold value, the probability of falling down event may be higher than ly-ing down event.

Table 1

k-fold cross-validation result.

Testing : Training k = 3 (%) k = 5 (%) k = 7 (%) k = 10 (%) k = 13 (%) k = 15 (%)

1:9 96.67 95.33 94.00 93.33 92.67 92.67

1:4 94.67 94.00 94.00 93.33 93.33 93.33

1:1 94.67 95.33 95.33 93.33 90.67 90.67

Average 95.34 94.89 94.44 93.33 92.22 92.22

(6)

In order to obtain more objective threshold value, 30 fall inci-dent cases are simulated and the transition time of the states are listed inTable 2. As shown inTable 2, the minimum transition time is 0.14 s, and the maximum is 0.34 s.Table 3shows the statistical result of the simulation.Table 3shows that the time difference is between 0.2262 s and 0.2658 s in 95% Conﬁdence Interval (CI). In other words, the time difference average will fall on this interval with 95% probability. Therefore, it is reasonable to set the upper bound of threshold value as 0.3 s. As mentioned above, the surveil-lance equipment in the system is a cheap web camera and it could process ﬁve frames in one second. As a result, the upper bound of the threshold value is 0.4 s in this system.

It is necessary to verify the above result, and statistical analysis is used to verify the result. Two statistical analyses will be con-ducted in this paper: check sample data distribution and deter-mine the threshold value. Hypothesis testing approach is adopted to verify the result, and the null hypothesis HO and alternative

hypothesis HAare described as follows:

HO: there is no difference between the distribution of the data

set and a normal one

HA: there is a difference between the distribution of the data set

and a normal one

In sample data distribution analysis, Kolmogorov–Smirnov test is conducted andTable 4is Kolmogorov–Smirnov test statements

provided by SPSS. The P-value inTable 4equals 0.749 which is lar-ger than 0.05, so we could not reject HO.Fig. 11is a

Quantile–Quan-tile plot which is constructed according to the simulation data in

Table 2.Fig. 11shows that the distribution of the data approaches to a straight-line, so this sample data can subject to normal distribution.

As described above, the assumption is that when fall incident happens, the average transition time of human body posture event will be less than 0.4 s and the transition time will subject to normal distribution. Statistical hypothesis testing is conducted to verify the critical value, and the null hypothesis HO and alternative

hypothesis HAare described as follows:

HO: Critical value

l

= 0.4

HA: Critical value

l

< 0.4

T-test is used to verify the assumption, andTable 5is the single sample T-test statistical statement andTable 6is the result of

sin-Table 2

Fall incident simulation result.

ID Temp state time Lying state time Difference

1 5.17 5.31 0.14 2 15.1 15.34 0.24 3 24.25 24.5 0.25 4 34.1 34.32 0.22 5 44.25 44.5 0.25 6 1.93 2.15 0.22 7 7.4 7.7 0.3 8 13.57 13.79 0.22 9 19.87 20.1 0.23 10 26.29 26.5 0.21 11 2.89 3.06 0.17 12 7.14 7.36 0.22 13 11.26 11.51 0.25 14 15.84 16.18 0.34 15 22.34 22.61 0.27 16 2.36 2.61 0.25 17 12.74 13.05 0.31 18 25.28 25.57 0.29 19 33.1 33.45 0.35 20 47.62 47.86 0.24 21 2.55 2.81 0.26 22 14.98 15.17 0.19 23 25.5 25.67 0.17 24 31.17 31.37 0.2 25 39.94 40.21 0.27 26 3.96 4.21 0.25 27 11.25 11.56 0.31 28 26.33 26.51 0.18 29 36.12 36.47 0.35 30 45.89 46.12 0.23 Table 3

Statistical information about fall incidents simulation.

Result Standard error mean

Difference Mean .2460 .00970 95% CI of difference Lower bound .2262 Upper bound .2658 5% Trimmed mean .2456 Table 4 Kolmogorov–Smirnov test. Difference N 30

Normal parameters Mean .2460

Std. deviation .05315 Most extreme differences Absolute .137

Positive .137

Negative .079

Kolmogorov–Smirnov Z .749

Asymp. Sig. (two-tailed) .630

Fig. 11. Normal Q–Q plot of difference.

Table 5

One-sample statistics.

N Mean Std. deviation Std. error mean

Difference 30 .2460 .05315 .00970

Table 6 One-sample test.

Test value = 0.4

95% CI t df Sig. (two-tailed) Mean diff. Lower Upper Diff. 15.870 29 .000 .15400 .1738 .1342

(7)

gle sample T-test. As shown inTable 6, the P-value in 1-tailed is less than 0.05 and that means that HOwill be rejected at

a

= 0.05

signiﬁcance level. In other words, the alternative hypothesis HA

will be accepted and that means the average transition time from temporary posture state to lying posture state is less than 0.4 s when fall incident occurs. Thus, it is reasonable to use 0.4 s as the critical value to determine whether fall incident is taking place.

4. Experiments and results

Table 7shows the equipment list in the experiment. The system will focus on how to increase accuracy rate of fall incident detec-tion and discriminate fall incident and lying down events to reduce the number of false alarms.

4.1. Postures analysis

As described above, the human body postures are divided into standing posture, temporary posture, and lying posture,

respec-tively. In the experimental environment, a ﬁxed camera is used to monitor indoor activities when human body appears in the scene.Fig. 12shows that the system detects standing, temporary, and lying postures. Meanwhile, Fig. 13 shows that the system could identify the posture even though the user stretches his hand in various way to affect the ratio of the human body height and width.

4.2. Fall detection

In addition to the kNN classiﬁcation model for posture classiﬁ-cation, falling speed is used to determine whether a fall incident event occurs. Meanwhile, if the object’s state transits from tempo-rary posture state to lying down posture state in less than thresh-old value, the system will regard this transition as a fall incident.

Fig. 14(a) and (b) shows the standing event and temporary event, respectively, whileFig. 14(c) shows the result of detecting fall inci-dent event.

4.3. Result and discussion

In addition to the above experiments, 15 people including 10 males and 5 females are invited to join fall detection system exper-iment. Age of these people ranges from 24 to 60, weight ranges from 90 lb to 220 pounds and height ranges from 5 ft. to 6 ft. When

Fig. 12. Three different postures identiﬁcation.

Fig. 13. Upper limb activities simulation.

Fig. 14. Experiment result. Table 7

Hardware list in the experiment.

Web camera Logic QuickCam STX

CPU Intel(R) Core(TM)2 2.13 GHz

(8)

each person comes into the scene, he/she will try to simulate the lying down cases that they may encounter in daily life for three times and simulate fall incident cases that they may encounter for another three times. These simulations include different types of falls, and there are 45 lying down cases and 45 fall incident cases. Table 8 shows the result. The results in the experiment shows that the total accuracy rate is about 4.44%, the accuracy rate on lying down events is about 86.67%, and accuracy rate on fall incident detection is about 82.22%.

5. Conclusion

The main purpose of this paper is to use off-the-shelf devices to develop a fall detection system. In practice, video surveillance sys-tem may infringe privacy to those who use the syssys-tem, so human body silhouette is adopted to improve privacy protection. Mean-while, human body silhouette removed most of human body fea-tures, so height and width of human body silhouette are adopted as the features of fall incident detection in many studies. Practi-cally, people may stretch out their hand to fetch something and that may affect the posture detection. In this paper, vertical projec-tion histograms of the silhouette image and statistical method are performed on the projection to reduce the influence. In feature selection, the ratio of bounding box height and width and the dif-ference of bounding box height and width are adopted as the sys-tem features in this paper. In fall incident detection, the syssys-tem makes use of kNN to classify human body postures and the best k value in kNN algorithm is determined by k-fold cross-validation scheme. Meanwhile, a falling detection flow is adopted to detect fall incidents and the transition time from temporary posture state to lying down posture state is the critical factor used for fall inci-dent detection. Statistical analysis is used to find out the transition time from the experiment data, and the value is verified by statis-tical hypothesis testing. The experiments show that it could reduce the effect of upper limb activities and the system has a correct rate of 84.44% on fall detection and lying down event detection.

Acknowledgment

This work was supported in part by the National Science Coun-cil under the Grants NSC-97-2221-E-009-135 and NSC-97-2811-E-009-019.

References

Anderson, D., Keller, J. M., Skubic, M., Chen, X., & He, Z. (2006). Recognizing falls from silhouettes. In Proceedings of the 28th IEEE EMBS annual international conference.

Bramberger, M., Doblander, A., Maier, A., Rinner, B., & Schwabach, H. (2006). Distributed embedded smart cameras for surveillance applications. IEEE Computer Magazine, 39(2), 6875.

Bureau, U. C. (2008). Population proﬁle of the united states.<http://www.census. gov/population/www/pop-proﬁle/elderpop.html>.

Cucchiara, R., Grana, C., Prati, A., & Vezzani, R. (2005). Probabilistic posture classiﬁcation for human-behavior analysis. IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans, 35(1), 42–54.

Duda, R. O., Hart, P. E., & Stork, D. G. (2000). Pattern classiﬁcation. Wiley-Interscience. GS, S. (1988). Falls among the elderly: Epidemiology and prevention. American

Journal of Preventive Medicine, 4(5), 282–288.

Juang, C.-F., & Chang, C.-M. (2007). Human body posture classiﬁcation by a neural fuzzy network and home care system application. IEEE Transactions on Systems, Man, and Cybernetics Part A: Systems and Humans, 37(6), 984–994.

Miaou, S.-G., Sung, P.-H., & Huang, C.-Y. (2006). A customized human fall detection system using omni-camera images and personal information. In Proceedings of the ﬁrst distributed diagnosis and home healthcare (D2H2) conference (pp. 39–42). Nait-Charif, H., & McKenna, S. J. (2004). Activity summarisation and fall detection in a supportive home environment. In Proceedings of the 17th international conference on pattern recognition (Vol. 4, pp. 323–326).

Nasution, A. H., & Emmanuel, S. (2007). Intelligent video surveillance for monitoring elderly in home environments. In IEEE ninth workshop on multimedia signal processing (pp. 203–206).

Murphy, S. L. (2000). Final data for 1998 national vital statistics reports. Tech. rep., National Center for Health Statistics.

Noury, N., Herd, T., Rialle, V., Virone, G., Mercier, E., Morey, G., Moro, A., & Porcheron, T. (2000). Monitoring behavior in home using a smart fall sensor and position sensors. In Proceedings of the IEEE international conference on microtechnologies in medicine and biology (pp. 607–610).

Shapiro, L. G., & Stockman, G. C. (2001). Computer vision. Prentice Hall.

Sterling, D. A., O’Connor, J. A., & Bonadies, J. (2001). Geriatric falls: Injury severity is high and disproportionate to mechanism. Journal of Trauma-Injury Infection and Critical Care, 50(1), 116–119.

Tabar, A. M., Keshavarz, A., & Aghajan, H. (2006). Smart home care network using sensor fusion and distributed vision-based reasoning. In VSSN’06: Proceedings of the fourth ACM international workshop on video surveillance and sensor networks (pp. 145–154).

Tamura, T., Yoshimura, T., Horiuchi, F., Higashi, Y., & Fujimoto, T. (2000). An ambulatory fall monitor for the elderly. In Proceedings of the IEEE international conference on engineering in medicine and biology society (Vol. 4, pp. 2608–2610). Tao, J., Turjo, M., Wong, M.-F., Wang, M., & Tan, Y.-P. (2005). Fall incidents detection for intelligent video surveillance. In Proceedings of the 15th international conference on communications and signal processing (pp. 1590–1594). Table 8

Simulation result.

Fall accident system recognition Yes No

Positive 37 6