A Novel On-Wrist Fall Detection System Using SDL Technique
191
data to a high and sparse dimensional representation using a learned dictionary
D, to make a more discriminated pattern and easier to be distinguished. In this
paper, we consider three SDL algorithms, SRC, FDDL, and LRSDL, that we
previously detailed in Sect. 2.2.
Considering the test phase, the testing sample can be classified by directly
coding it over the obtained D. Generally, the sparse code is then used as a feature
descriptor of the data in order to calculate the reconstruction error associated to
each class. The prediction is accorded to the class with the least error following
the formula expressed by Eq. (4). However, the SDL performance is directly
affected by the dictionary size. To abstract each SDL’s higher performance we
will inspect the impact of the Dictionary size into the system’s accuracy.
4 Experimental Validation
4.1 Performance Metrics
This study is evaluated in terms of three common metrics, namely, Accuracy
(AC), Sensitivity (SE), and Specificity (SP). AC represents the overall true detection, SE represents the ability to detect authentic falls among all detected falls,
and SP represents the capacity to detect real ADL in all the detected ADL.
4.2 Experimental Configuration
In our experimental analysis, we assume that a 4-s time window is sufficient
to extract a fall or an ADL event. We consider that the collected data set is
subdivided such as 75% of the data (nearly 300 samples for each class) is for
the training phase and 25% for the test phase. From our experiments, the SDL
algorithms’ hyper-parameters are set based on the best-achieved performances
for our dataset using random training features. Thus, we initiate them as follows:
SRC: λ = 0.01; FDDL: λ 1 = λ 2 = 0.001; LRSDL: λ 1 = 0.001, λ 2 = 0.01,
η = 0.02. Throughout this study, the size of the dictionary D for the FDDL and
the LRSDL algorithm will vary between 50 and 300 atoms per class depending
on the experiment.
4.3 Experimental Result
In this study, we followed two main experimentation scheme to validate the high
sensitivity and efficiency of our proposed method. Firstly, we fix the Dictionary
size in order to assess the SDL classification performance behavior compared
with each outline of both scenarios showcased in Fig. 2. Secondly, we evaluate
the best performance of the previous experiment with multiple D sizes for the
FDDL and LRSDL algorithms to exhibit for each the best-fitted size to our
proposed system.
191
data to a high and sparse dimensional representation using a learned dictionary
D, to make a more discriminated pattern and easier to be distinguished. In this
paper, we consider three SDL algorithms, SRC, FDDL, and LRSDL, that we
previously detailed in Sect. 2.2.
Considering the test phase, the testing sample can be classified by directly
coding it over the obtained D. Generally, the sparse code is then used as a feature
descriptor of the data in order to calculate the reconstruction error associated to
each class. The prediction is accorded to the class with the least error following
the formula expressed by Eq. (4). However, the SDL performance is directly
affected by the dictionary size. To abstract each SDL’s higher performance we
will inspect the impact of the Dictionary size into the system’s accuracy.
4 Experimental Validation
4.1 Performance Metrics
This study is evaluated in terms of three common metrics, namely, Accuracy
(AC), Sensitivity (SE), and Specificity (SP). AC represents the overall true detection, SE represents the ability to detect authentic falls among all detected falls,
and SP represents the capacity to detect real ADL in all the detected ADL.
4.2 Experimental Configuration
In our experimental analysis, we assume that a 4-s time window is sufficient
to extract a fall or an ADL event. We consider that the collected data set is
subdivided such as 75% of the data (nearly 300 samples for each class) is for
the training phase and 25% for the test phase. From our experiments, the SDL
algorithms’ hyper-parameters are set based on the best-achieved performances
for our dataset using random training features. Thus, we initiate them as follows:
SRC: λ = 0.01; FDDL: λ 1 = λ 2 = 0.001; LRSDL: λ 1 = 0.001, λ 2 = 0.01,
η = 0.02. Throughout this study, the size of the dictionary D for the FDDL and
the LRSDL algorithm will vary between 50 and 300 atoms per class depending
on the experiment.
4.3 Experimental Result
In this study, we followed two main experimentation scheme to validate the high
sensitivity and efficiency of our proposed method. Firstly, we fix the Dictionary
size in order to assess the SDL classification performance behavior compared
with each outline of both scenarios showcased in Fig. 2. Secondly, we evaluate
the best performance of the previous experiment with multiple D sizes for the
FDDL and LRSDL algorithms to exhibit for each the best-fitted size to our
proposed system.