Human falls are a major health issue for the elderly. Falls occur in 13% of the population over the age of 65, and 13% of this group lives alone. Falls and instability are common causes of death
The fall detection market is divided into three segments. It can be divided into wearable and non-wearable sensor segments based on the various types of sensors. Wearable sensors are easy to use, lightweight, and easily portable from one place to other whereas non-wearable sensors are fitted in the room, corridor, or other outdoor areas. If we consider technology, then the system market can be divided into GPS systems, mobile phones, and sensor-based market segments. By considering end-user fall detection, the market can be segmented into home care and senior assisted living facilities.
There are many important factors of fall detection
The human fall detector helps to reduce the time of lying on the floor after falling. This is the key factor to reduce the severity of falls.
Fall detection reduces the fear of falling in the elders. They have confidence that if they fall unconsciously anytime then a message of their falling will be immediately sent to their caretaker.
Fall detection reduces medical expenses because fall detection reduces the severity of falls.
Fall detection can also be implemented on the Internet of Things (IoT) environment as shown in
Fall detection methods can be divided into various categories based on the different types of technologies used. According to
Wearable sensors based fall detection
Ambient sensors based fall detection
Vision-based fall detection
Various categories of human fall detection methods have been shown in
Binary classification is a classification method that divides data into two groups: positive and negative. The scenario of positive and negative classes gives rise to the concept of one-class classification. In a one-class classification, Generally data of negative class is available in very less amount. As a result, the problem of one-class classification is more difficult than the problem of conventional multi-class / binary classification. Therefore, one-class classification algorithms should be expected to require a greater number of training instances than traditional multi-class classification algorithms
One class classification method is divided into two approaches the first approach is based on the availability of training data and the second approach is based on the type of algorithm used as shown in
It is relevant to the case where we have to classify whether the machine, operating normally or not. In this case, the dataset of the normal behavior of the machine is available but fault occurs very less so data of fall activities are available in very fewer numbers. Most of the time dataset of abnormal behavior of the machine is not available. In this type of case, one class classifier classifies the normal behavior of machines in one class, and behavior other than normal behavior is considered as outlier. One class classification is also known as outlier detection, novelty detection. The term one class classifier was firstly coined by Minter (1975). In a Bayes classification learning context he had been the first to use the term 'one class classification' four decades ago and only labeled 'class of interest' data is required. After that Moya et al.
We have used one-class classification in our proposed method of fall detection because, in real-life scenarios, data of normal daily activities are easily available but data of fall activities are rarely available because fall activities in elder persons are normally not recorded or not available. Two different predictions are possible in one classification, which means an instance belongs to the studied class during training and unknown in which the instance does not seem to be part of the previous class
To analyze the sensor signal pattern of human fall detection, much research works have been carried out. In 2017
Our method is based on machine learning but used one-class classifier instead of a multiclass or binary classifier. In various applications such as smart home
Some fall detection methods
In the proposed method bagging-based one-class classification for fall detection has been used. In real life real fall detection data is not available and data of normal daily activities are easily available and can be easily recorded. This is the reason behind one class classification-based method. This method considers daily living activities as targets and activities other than daily activities are considered as outlier i.e. fall activities.
Thus we got a vector of 51 values of accelerometer resultant
Now we have a vector of 55 values for each ADL or fall activity vector. We have analyzed each vector of ADL or fall activity containing 55 values using one-class classification approach
In this experiment, data of only ADL activities have been used to train the classifier, and ADL and fall activities have been used to test the classifier. ADL activities are considered as target and fall activities as considered as outlier. When we test the model, ADL activities are classified as targets, and fall activities are considered as unknown instances.
The proposed model is tested on both datasets. In Medrano et. al.
We test the model in both types of data i.e. ADL activities and fall activities. We also test the model in the Mobifall dataset containing eight types of falls. Following are the results when we test the model on both datasets i.e. Medrano et. al.
S.No |
Performance measure |
When a classification model is built without features |
When a classification model is built with features |
||
When smart phone is in pocket |
When smart phone is in handbag |
When smart phone is in pocket |
When smart phone is in handbag |
||
1 |
Sensitivity |
100.0000 |
100.0000 |
100.0000 |
100.0000 |
2 |
Specificity |
88.8690 |
90.7112 |
94.9207 |
93.6139 |
3 |
Accuracy |
89.5420 |
92.5869 |
95.2278 |
94.9035 |
4 |
Precision |
36.6351 |
73.1469 |
55.8889 |
79.8473 |
5 |
F1 score |
53.6247 |
84.4911 |
71.7035 |
88.7946 |
6 |
G Mean |
94.2703 |
95.2424 |
97.4272 |
96.7543 |
7 |
Critical Success Index(CSI) |
36.6351 |
73.1469 |
55.8889 |
79.8473 |
8 |
Matthews correlation coefficient (MCC) |
57.0590 |
81.4570 |
72.8355 |
86.4571 |
9 |
Bookmaker Informedness (BM) |
187.8690 |
189.7112 |
193.9207 |
192.6139 |
S.No |
Activity |
Fall prediction accuracy |
|
Bagging based one-class classification on the dataset without features. |
Bagging based one-class classification on the dataset with features. |
||
1 |
Forward lying (FOL) fall |
100% |
100% |
2 |
Front knee lying (FKL) fall |
100% |
100% |
3 |
Back sitting chair (BSC) fall |
100% |
100% |
4 |
Sideward lying (SDL) fall |
100% |
100% |
From
From
We have also studied the effect of each feature on the performance of one class classification model as shown in
S.No |
Performance measure |
Features used to build the model Std, RMS, En |
Features used to build the model Var, RMS, En |
||
When smartphone is in pocket |
When smartphone is in handbag |
When smartphone is in pocket |
When smartphone is in handbag |
||
1 |
Sensitivity |
100.0000 |
100.0000 |
100.0000 |
100.0000 |
2 |
Specificity |
88.5235 |
90.1306 |
94.4857 |
93.3237 |
3 |
Accuracy |
89.2175 |
92.1236 |
94.8191 |
94.6718 |
4 |
Precision |
35.9286 |
71.9395 |
53.8544 |
79.1225 |
5 |
F1 score |
52.8639 |
83.6800 |
70.0070 |
88.3446 |
6 |
G Mean |
94.0869 |
94.9371 |
97.2037 |
96.6042 |
7 |
Critical Success Index(CSI) |
35.9286 |
71.9395 |
53.8544 |
79.1225 |
8 |
Matthews correlation coefficient (MCC) |
56.3961 |
80.5230 |
71.3335 |
85.9302 |
9 |
Bookmaker Informedness (BM) |
187.5235 |
189.1306 |
193.4857 |
192.3237 |
S.No |
Performance measure |
Features used to build the model Std, Var, En |
Features used to build the model Std, Var, RMS |
||
When smart phone is in pocket |
When smart phone is in handbag |
When smart phone is in pocket |
When smart phone is in handbag |
||
1 |
Sensitivity |
100.0000 |
100.0000 |
100.0000 |
100.0000 |
2 |
Specificity |
94.9335 |
93.6139 |
94.9335 |
93.6139 |
3 |
Accuracy |
95.2398 |
94.9035 |
95.2398 |
94.9035 |
4 |
Precision |
55.9511 |
79.8473 |
55.9511 |
79.8473 |
5 |
F1 score |
71.7546 |
88.7946 |
71.7546 |
88.7946 |
6 |
G Mean |
97.4338 |
96.7543 |
97.4338 |
96.7543 |
7 |
Critical Success Index(CSI) |
55.9511 |
79.8473 |
55.9511 |
79.8473 |
8 |
Matthews correlation coefficient (MCC) |
72.8809 |
86.4571 |
72.8809 |
86.4571 |
9 |
Bookmaker Informedness (BM) |
193.9335 |
192.6139 |
193.9335 |
192.6139 |
From
We have also compared the classification of the dataset using one-class classification and one-class support vector machine classification
S.No. |
Measures for comparison for one-class classification |
Bagging classifier |
SVM classifier |
1 |
Number of training instances |
8894.70 |
8894.70 |
2 |
Number of Testing instances |
988.30 |
988.30 |
3 |
Percentage of correctly classified instances |
88.27 |
11.51 |
4 |
Percentage of incorrectly classified instances |
0 |
0 |
5 |
Percentage of unclassified instances |
11.73 |
88.49 |
6 |
True positive |
1 |
1 |
7 |
False negative |
0 |
0 |
8 |
Time for training |
159.96 |
131.25 |
9 |
Time for testing |
0.01 |
3.41 |
From the results, it can be seen that percentage of correctly classified instances are 88.27% and the percentage of incorrectly classified instances are 11.73 % while building the model using one-class classification based on bagging classifier and correctly classified instances are 11.57% and the percentage of incorrectly classified instances are 88.49% while building the model using a one-class classification based on support vector machine. From
Human falls can be detected based on the approach that its accelerometer data during fall is different than accelerometer data of normal living activities. Two publically available fall detection datasets were used in this method. The PKI descritize filter has been applied for filtering accelerometer data. Five different features were calculated on that dataset for classification. After that one class classification was used to build the model because real datasets of fall activities are not available, all fall datasets are simulated in lab. It is found that that proposed method gave very good results on both datasets i.e. all ADL activities were correctly classified as the target and all fall activities were correctly classified as an outlier. This one-class classification is very important because it can be effectively applied in real-life fall. In this proposed method, it is also found that variance is an important feature while building a fall detection model. Although this work is tested on a limited number of the dataset but in the future, this work can also be extended by incorporating more datasets in the study to validate its reliability on a large population.